In [2]:
%autosave 0
Autosave disabled

Artificial Intelligence Nanodegree

Convolutional Neural Networks

Project: Write an Algorithm for a Dog Identification App


In this notebook, some template code has already been provided for you, and you will need to implement additional functionality to successfully complete this project. You will not need to modify the included code beyond what is requested. Sections that begin with '(IMPLEMENTATION)' in the header indicate that the following block of code will require additional functionality which you must provide. Instructions will be provided for each section, and the specifics of the implementation are marked in the code block with a 'TODO' statement. Please be sure to read the instructions carefully!

Note: Once you have completed all of the code implementations, you need to finalize your work by exporting the iPython Notebook as an HTML document. Before exporting the notebook to html, all of the code cells need to have been run so that reviewers can see the final implementation and output. You can then export the notebook by using the menu above and navigating to \n", "File -> Download as -> HTML (.html). Include the finished document along with this notebook as your submission.

In addition to implementing code, there will be questions that you must answer which relate to the project and your implementation. Each section where you will answer a question is preceded by a 'Question X' header. Carefully read each question and provide thorough answers in the following text boxes that begin with 'Answer:'. Your project submission will be evaluated based on your answers to each of the questions and the implementation you provide.

Note: Code and Markdown cells can be executed using the Shift + Enter keyboard shortcut. Markdown cells can be edited by double-clicking the cell to enter edit mode.

The rubric contains optional "Stand Out Suggestions" for enhancing the project beyond the minimum requirements. If you decide to pursue the "Stand Out Suggestions", you should include the code in this IPython notebook.


Why We're Here

In this notebook, you will make the first steps towards developing an algorithm that could be used as part of a mobile or web app. At the end of this project, your code will accept any user-supplied image as input. If a dog is detected in the image, it will provide an estimate of the dog's breed. If a human is detected, it will provide an estimate of the dog breed that is most resembling. The image below displays potential sample output of your finished project (... but we expect that each student's algorithm will behave differently!).

Sample Dog Output

In this real-world setting, you will need to piece together a series of models to perform different tasks; for instance, the algorithm that detects humans in an image will be different from the CNN that infers dog breed. There are many points of possible failure, and no perfect algorithm exists. Your imperfect solution will nonetheless create a fun user experience!

The Road Ahead

We break the notebook into separate steps. Feel free to use the links below to navigate the notebook.

  • Step 0: Import Datasets
  • Step 1: Detect Humans
  • Step 2: Detect Dogs
  • Step 3: Create a CNN to Classify Dog Breeds (from Scratch)
  • Step 4: Use a CNN to Classify Dog Breeds (using Transfer Learning)
  • Step 5: Create a CNN to Classify Dog Breeds (using Transfer Learning)
  • Step 6: Write your Algorithm
  • Step 7: Test Your Algorithm

Step 0: Import Datasets

Import Dog Dataset

In the code cell below, we import a dataset of dog images. We populate a few variables through the use of the load_files function from the scikit-learn library:

  • train_files, valid_files, test_files - numpy arrays containing file paths to images
  • train_targets, valid_targets, test_targets - numpy arrays containing onehot-encoded classification labels
  • dog_names - list of string-valued dog breed names for translating labels
In [3]:
from sklearn.datasets import load_files       
from keras.utils import np_utils
import numpy as np
from glob import glob

# define function to load train, test, and validation datasets
def load_dataset(path):
    data = load_files(path)
    dog_files = np.array(data['filenames'])
    dog_targets = np_utils.to_categorical(np.array(data['target']), 133)
    return dog_files, dog_targets

# load train, test, and validation datasets
train_files, train_targets = load_dataset('dogImages/train')
valid_files, valid_targets = load_dataset('dogImages/valid')
test_files, test_targets = load_dataset('dogImages/test')

# load list of dog names
dog_names = [item[20:-1] for item in sorted(glob("dogImages/train/*/"))]

# print statistics about the dataset
print('There are %d total dog categories.' % len(dog_names))
print('There are %s total dog images.\n' % len(np.hstack([train_files, valid_files, test_files])))
print('There are %d training dog images.' % len(train_files))
print('There are %d validation dog images.' % len(valid_files))
print('There are %d test dog images.'% len(test_files))
Using TensorFlow backend.
There are 133 total dog categories.
There are 8351 total dog images.

There are 6680 training dog images.
There are 835 validation dog images.
There are 836 test dog images.
In [4]:
print(train_files[2], '\n', train_targets[2])
dogImages/train/088.Irish_water_spaniel/Irish_water_spaniel_06014.jpg 
 [ 0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.
  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.
  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.
  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.
  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  1.  0.  0.
  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.
  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.
  0.  0.  0.  0.  0.  0.  0.]

Import Human Dataset

In the code cell below, we import a dataset of human images, where the file paths are stored in the numpy array human_files.

In [5]:
import random
random.seed(853092)

# load filenames in shuffled human dataset
human_files = np.array(glob("lfw/*/*"))
random.shuffle(human_files)

# print statistics about the dataset
print('There are %d total human images.' % len(human_files))
There are 13229 total human images.

Step 1: Detect Humans

We use OpenCV's implementation of Haar feature-based cascade classifiers to detect human faces in images. OpenCV provides many pre-trained face detectors, stored as XML files on github. We have downloaded one of these detectors and stored it in the haarcascades directory.

In the next code cell, we demonstrate how to use this detector to find human faces in a sample image.

In [6]:
import cv2
In [7]:
import cv2                
import matplotlib.pyplot as plt                        
%matplotlib inline                               

# extract pre-trained face detector
face_cascade = cv2.CascadeClassifier('haarcascades/haarcascade_frontalface_alt.xml')

# load color (BGR) image
img = cv2.imread(human_files[19])
# convert BGR image to grayscale
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)

# find faces in image
faces = face_cascade.detectMultiScale(gray)

# print number of faces detected in the image
print('Number of faces detected:', len(faces))

# get bounding box for each detected face
for (x,y,w,h) in faces:
    # add bounding box to color image
    cv2.rectangle(img,(x,y),(x+w,y+h),(255,0,0),2)
    
# convert BGR image to RGB for plotting
cv_rgb = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)

# display the image, along with bounding box
plt.imshow(cv_rgb)
plt.show()

#print ('imagedata\n',gray.shape)
Number of faces detected: 1

Before using any of the face detectors, it is standard procedure to convert the images to grayscale. The detectMultiScale function executes the classifier stored in face_cascade and takes the grayscale image as a parameter.

In the above code, faces is a numpy array of detected faces, where each row corresponds to a detected face. Each detected face is a 1D array with four entries that specifies the bounding box of the detected face. The first two entries in the array (extracted in the above code as x and y) specify the horizontal and vertical positions of the top left corner of the bounding box. The last two entries in the array (extracted here as w and h) specify the width and height of the box.

Write a Human Face Detector

We can use this procedure to write a function that returns True if a human face is detected in an image and False otherwise. This function, aptly named face_detector, takes a string-valued file path to an image as input and appears in the code block below.

In [8]:
# returns "True" if face is detected in image stored at img_path
def face_detector(img_path):
    img = cv2.imread(img_path)
    gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
    faces = face_cascade.detectMultiScale(gray)
    if len(faces) == 0 and img_path.startswith('lfw'):
        print(img_path)
    return len(faces) > 0

(IMPLEMENTATION) Assess the Human Face Detector

Question 1: Use the code cell below to test the performance of the face_detector function.

  • What percentage of the first 100 images in human_files have a detected human face?
  • What percentage of the first 100 images in dog_files have a detected human face?

Ideally, we would like 100% of human images with a detected face and 0% of dog images with a detected face. You will see that our algorithm falls short of this goal, but still gives acceptable performance. We extract the file paths for the first 100 images from each of the datasets and store them in the numpy arrays human_files_short and dog_files_short.

Answer:

  • What percentage of the first 100 images in human_files have a detected human face?
    The human face detector works accurately 99% of the time, misses 1 face (idx=6, as seen above).
  • What percentage of the first 100 images in dog_files have a detected human face?
    11% of the dog faces are incorrectly identified as human faces.
In [9]:
human_files_short = human_files[:100]
dog_files_short = train_files[:100]
# Do NOT modify the code above this line.

#print (face_detector(dog_files_short[10]))
## TODO: Test the performance of the face_detector algorithm 
## on the images in human_files_short and dog_files_short.

h_det = list(map(face_detector, human_files_short))
d_det = list(map(face_detector, dog_files_short))
lfw/Hernan_Diaz/Hernan_Diaz_0001.jpg
In [10]:
print ('{}\n{}'.format(sum(h_det),sum(d_det)))
99
11

Question 2: This algorithmic choice necessitates that we communicate to the user that we accept human images only when they provide a clear view of a face (otherwise, we risk having unneccessarily frustrated users!). In your opinion, is this a reasonable expectation to pose on the user? If not, can you think of a way to detect humans in images that does not necessitate an image with a clearly presented face?

Answer:

In the real world, this is not a reasonable request. A human can recognize a face even without a clear view of the face, e.g. sideways, tilted, or even upside-down.
For our checkimg, let's try out an mlp first, and then the CNN (using models as-is from the aind-cnn project)

We suggest the face detector from OpenCV as a potential way to detect human images in your algorithm, but you are free to explore other approaches, especially approaches that make use of deep learning :). Please use the code cell below to design and test your own face detection algorithm. If you decide to pursue this optional task, report performance on each of the datasets.

Reference: https://hjweide.github.io/efficient-image-loading

In [11]:
## (Optional) TODO: Report the performance of another  
## face detection algorithm on the LFW dataset
### Feel free to use as many code cells as needed.
In [12]:
# First the MLP

# We need to load the human faces into x_train, y_train, x_test, y_test
#cv2.CV_LOAD_IMAGE_GRAYSCALE has an integer value of 0. use that in imread

def getX(image_dir):
    data = np.empty((len(image_dir), 62500), dtype=np.uint8)
    labels = np.empty((len(image_dir), 2), dtype=np.uint8)
    for i, fpath in enumerate(image_dir):
        #cv2.CV_LOAD_IMAGE_GRAYSCALE has an integer value of 0. use that in imread
        img = cv2.imread(fpath, 0)
        if img.shape is not (250, 250):
            img = cv2.resize(img, (250,250))
        img = img.flatten()
        data[i, ...] = img
        labels[i] = [1,0] if 'lfw' in fpath else [0, 1]
        
    return data,labels

def load_human_and_dog_faces(tr=100, vl=25, ts=50):
    tr_data, tr_labels = getX(np.concatenate([human_files[:tr], train_files[:tr]]))
    vl_data, vl_labels = getX(np.concatenate([human_files[tr:tr+vl], train_files[tr:tr+vl]]))
    ts_data, ts_labels = getX(np.concatenate([human_files[tr+vl:tr+vl+ts], train_files[tr+vl:tr+vl+ts]]))
    return tr_data, tr_labels, vl_data, vl_labels, ts_data, ts_labels


x_train, y_train, x_valid, y_valid, x_test, y_test = load_human_and_dog_faces(tr=500, vl=50, ts=50)

print(x_train.shape, y_train.shape, x_valid.shape, y_valid.shape, x_test.shape, y_test.shape)

num_classes = 2
from keras.models import Sequential
from keras.layers import Dense, Dropout, Flatten

# define the model
model = Sequential()
model.add(Dense(6250, input_shape=(62500,), activation='relu'))
model.add(Dense(512, activation='relu'))
model.add(Dropout(0.8))
model.add(Dense(512, activation='relu'))
model.add(Dropout(0.1))
model.add(Dense(num_classes, activation='softmax'))

# compile the model
model.compile(loss='categorical_crossentropy', optimizer='rmsprop', 
                  metrics=['accuracy'])

run_fit = False #set to True when ready to run fit()
#model.summary()

from keras.callbacks import ModelCheckpoint   

if run_fit:
    # train the model, and save the best 
    checkpointer = ModelCheckpoint(filepath='saved_models/MLP.human.weights.best.hdf5', verbose=1, 
                               save_best_only=True)
    hist = model.fit(x_train, y_train, batch_size=32, epochs=2,
          validation_data=(x_valid, y_valid), callbacks=[checkpointer], 
          verbose=2, shuffle=True)
    # load the weights that yielded the best validation accuracy
    model.load_weights('saved_models/MLP.human.weights.best.hdf5')
    # evaluate and print test accuracy
    score = model.evaluate(x_test, y_test, verbose=0)
    print('\n', 'Test accuracy:', score[1])
(1000, 62500) (1000, 2) (100, 62500) (100, 2) (100, 62500) (100, 2)
(1000, 62500) (1000, 2) (100, 62500) (100, 2) (100, 62500) (100, 2)
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
dense_1 (Dense)              (None, 6250)              390631250 
_________________________________________________________________
dense_2 (Dense)              (None, 512)               3200512   
_________________________________________________________________
dropout_1 (Dropout)          (None, 512)               0         
_________________________________________________________________
dense_3 (Dense)              (None, 512)               262656    
_________________________________________________________________
dropout_2 (Dropout)          (None, 512)               0         
_________________________________________________________________
dense_4 (Dense)              (None, 2)                 1026      
=================================================================
Total params: 394,095,444.0
Trainable params: 394,095,444.0
Non-trainable params: 0.0
_________________________________________________________________
Train on 1000 samples, validate on 100 samples
Epoch 1/2
Epoch 00000: val_loss improved from inf to 8.05905, saving model to MLP.human.weights.best.hdf5
650s - loss: 8.1076 - acc: 0.4970 - val_loss: 8.0590 - val_acc: 0.5000
Epoch 2/2
Epoch 00001: val_loss did not improve
186s - loss: 8.0752 - acc: 0.4990 - val_loss: 8.0590 - val_acc: 0.5000

 Test accuracy: 0.5
 
In [13]:
# Then the Conv2D

#Load the images in as color 3D because Conv2D can handle multi-dim inputs while perceptron needs a vector

def getX(image_dir):
    data = np.empty((len(image_dir), 250, 250, 3), dtype=np.uint8)
    labels = np.empty((len(image_dir), 2), dtype=np.uint8)
    for i, fpath in enumerate(image_dir):
        img = cv2.imread(fpath)
        if img.shape[0] is not 250 or img.shape[1] is not 250:
            img = cv2.resize(img, (250,250))
        #img = img.transpose(2, 0, 1)
        data[i, ...] = img
        labels[i] = [1,0] if 'lfw' in fpath else [0, 1]
        
    return data,labels

def load_human_and_dog_faces(tr=100, vl=25, ts=50):
    tr_data, tr_labels = getX(np.concatenate([human_files[:tr], train_files[:tr]]))
    vl_data, vl_labels = getX(np.concatenate([human_files[tr:tr+vl], train_files[tr:tr+vl]]))
    ts_data, ts_labels = getX(np.concatenate([human_files[tr+vl:tr+vl+ts], train_files[tr+vl:tr+vl+ts]]))
    return tr_data, tr_labels, vl_data, vl_labels, ts_data, ts_labels


x_train, y_train, x_valid, y_valid, x_test, y_test = load_human_and_dog_faces(tr=500, vl=50, ts=50)

#testpath='lfw/Harrison_Ford/Harrison_Ford_0003.jpg'
#testpath = 'dogImages/train/001.Affenpinscher/Affenpinscher_00001.jpg'
#print(testpath)
#img = cv2.imread(testpath)
#print(img.shape)
#print(img.shape[0] is not 250 and img.shape[1] is not 250)
#img = cv2.resize(img,(250,250))
#print(img.shape)
#img = img.transpose(2, 0, 1)
#print(img.shape)
#print(cv2.resize(img,(250,250)).shape)
#print(img)
In [14]:
print(x_train.shape, y_train.shape)
(1000, 250, 250, 3) (1000, 2)
In [15]:
print(x_train.shape, y_train.shape)

from keras.models import Sequential
from keras.layers import Conv2D, MaxPooling2D, Flatten, Dense, Dropout

model = Sequential()
model.add(Conv2D(filters=16, kernel_size=2, padding='same', activation='relu', 
                        input_shape=(250, 250, 3)))
model.add(MaxPooling2D(pool_size=2))
model.add(Conv2D(filters=32, kernel_size=2, padding='same', activation='relu'))
model.add(MaxPooling2D(pool_size=2))
model.add(Conv2D(filters=64, kernel_size=2, padding='same', activation='relu'))
model.add(MaxPooling2D(pool_size=2))
model.add(Dropout(0.3))
model.add(Flatten())
model.add(Dense(500, activation='relu'))
model.add(Dropout(0.4))
model.add(Dense(2, activation='softmax'))

model.summary()

# compile the model
model.compile(loss='categorical_crossentropy', optimizer='rmsprop', 
                  metrics=['accuracy'])
(1000, 250, 250, 3) (1000, 2)
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
conv2d_1 (Conv2D)            (None, 250, 250, 16)      208       
_________________________________________________________________
max_pooling2d_1 (MaxPooling2 (None, 125, 125, 16)      0         
_________________________________________________________________
conv2d_2 (Conv2D)            (None, 125, 125, 32)      2080      
_________________________________________________________________
max_pooling2d_2 (MaxPooling2 (None, 62, 62, 32)        0         
_________________________________________________________________
conv2d_3 (Conv2D)            (None, 62, 62, 64)        8256      
_________________________________________________________________
max_pooling2d_3 (MaxPooling2 (None, 31, 31, 64)        0         
_________________________________________________________________
dropout_3 (Dropout)          (None, 31, 31, 64)        0         
_________________________________________________________________
flatten_1 (Flatten)          (None, 61504)             0         
_________________________________________________________________
dense_5 (Dense)              (None, 500)               30752500  
_________________________________________________________________
dropout_4 (Dropout)          (None, 500)               0         
_________________________________________________________________
dense_6 (Dense)              (None, 2)                 1002      
=================================================================
Total params: 30,764,046.0
Trainable params: 30,764,046.0
Non-trainable params: 0.0
_________________________________________________________________
In [16]:
from keras.preprocessing.image import ImageDataGenerator

# create and configure augmented image generator
datagen_train = ImageDataGenerator(
    width_shift_range=0.1,  # randomly shift images horizontally (10% of total width)
    height_shift_range=0.1,  # randomly shift images vertically (10% of total height)
    horizontal_flip=True) # randomly flip images horizontally

# create and configure augmented image generator
datagen_valid = ImageDataGenerator(
    width_shift_range=0.1,  # randomly shift images horizontally (10% of total width)
    height_shift_range=0.1,  # randomly shift images vertically (10% of total height)
    horizontal_flip=True) # randomly flip images horizontally

# fit augmented image generator on data
datagen_train.fit(x_train)
datagen_valid.fit(x_valid)
In [16]:
# DO NOT re-run this cell. Takes a very long time. Output is described in the cell below.
from keras.callbacks import ModelCheckpoint   

batch_size = 32
epochs = 50

# train the model
checkpointer = ModelCheckpoint(filepath='saved_models/aug_model.human.weights.best.hdf5', verbose=1, 
                               save_best_only=True)
model.fit_generator(datagen_train.flow(x_train, y_train, batch_size=batch_size),
                    steps_per_epoch=x_train.shape[0] // batch_size,
                    epochs=epochs, verbose=2, callbacks=[checkpointer],
                    validation_data=datagen_valid.flow(x_valid, y_valid, batch_size=batch_size),
                    validation_steps=x_valid.shape[0] // batch_size)
Epoch 1/50
Epoch 00000: val_loss improved from inf to 8.22694, saving model to saved_models/aug_model.human.weights.best.hdf5
49s - loss: 7.8977 - acc: 0.5071 - val_loss: 8.2269 - val_acc: 0.4896
Epoch 2/50
Epoch 00001: val_loss improved from 8.22694 to 8.05905, saving model to saved_models/aug_model.human.weights.best.hdf5
45s - loss: 8.2348 - acc: 0.4891 - val_loss: 8.0590 - val_acc: 0.5000
Epoch 3/50
Epoch 00002: val_loss improved from 8.05905 to 7.82202, saving model to saved_models/aug_model.human.weights.best.hdf5
45s - loss: 7.8670 - acc: 0.5119 - val_loss: 7.8220 - val_acc: 0.5147
Epoch 4/50
Epoch 00003: val_loss did not improve
44s - loss: 7.9793 - acc: 0.5049 - val_loss: 8.0590 - val_acc: 0.5000
Epoch 5/50
Epoch 00004: val_loss did not improve
44s - loss: 8.0134 - acc: 0.5028 - val_loss: 8.5331 - val_acc: 0.4706
Epoch 6/50
Epoch 00005: val_loss improved from 7.82202 to 7.82202, saving model to saved_models/aug_model.human.weights.best.hdf5
45s - loss: 8.0753 - acc: 0.4990 - val_loss: 7.8220 - val_acc: 0.5147
Epoch 7/50
Epoch 00006: val_loss improved from 7.82202 to 7.58499, saving model to saved_models/aug_model.human.weights.best.hdf5
45s - loss: 7.9437 - acc: 0.5072 - val_loss: 7.5850 - val_acc: 0.5294
Epoch 8/50
Epoch 00007: val_loss did not improve
44s - loss: 8.3161 - acc: 0.4841 - val_loss: 8.0590 - val_acc: 0.5000
Epoch 9/50
Epoch 00008: val_loss did not improve
44s - loss: 7.9143 - acc: 0.5090 - val_loss: 8.5331 - val_acc: 0.4706
Epoch 10/50
Epoch 00009: val_loss did not improve
44s - loss: 7.9321 - acc: 0.5079 - val_loss: 8.0590 - val_acc: 0.5000
Epoch 11/50
Epoch 00010: val_loss did not improve
44s - loss: 8.0428 - acc: 0.5010 - val_loss: 7.8220 - val_acc: 0.5147
Epoch 12/50
Epoch 00011: val_loss did not improve
44s - loss: 8.2673 - acc: 0.4871 - val_loss: 7.8220 - val_acc: 0.5147
Epoch 13/50
Epoch 00012: val_loss did not improve
44s - loss: 8.2217 - acc: 0.4899 - val_loss: 7.8220 - val_acc: 0.5147
Epoch 14/50
Epoch 00013: val_loss did not improve
44s - loss: 8.0737 - acc: 0.4991 - val_loss: 7.5850 - val_acc: 0.5294
Epoch 15/50
Epoch 00014: val_loss did not improve
44s - loss: 8.0103 - acc: 0.5030 - val_loss: 8.0590 - val_acc: 0.5000
Epoch 16/50
Epoch 00015: val_loss improved from 7.58499 to 7.34796, saving model to saved_models/aug_model.human.weights.best.hdf5
45s - loss: 7.9127 - acc: 0.5091 - val_loss: 7.3480 - val_acc: 0.5441
Epoch 17/50
Epoch 00016: val_loss did not improve
44s - loss: 8.2364 - acc: 0.4890 - val_loss: 8.0590 - val_acc: 0.5000
Epoch 18/50
Epoch 00017: val_loss did not improve
45s - loss: 8.0250 - acc: 0.5021 - val_loss: 8.7701 - val_acc: 0.4559
Epoch 19/50
Epoch 00018: val_loss did not improve
45s - loss: 7.8817 - acc: 0.5110 - val_loss: 8.2961 - val_acc: 0.4853
Epoch 20/50
Epoch 00019: val_loss did not improve
45s - loss: 8.3517 - acc: 0.4818 - val_loss: 8.7701 - val_acc: 0.4559
Epoch 21/50
Epoch 00020: val_loss did not improve
46s - loss: 7.8980 - acc: 0.5100 - val_loss: 8.7701 - val_acc: 0.4559
Epoch 22/50
Epoch 00021: val_loss did not improve
46s - loss: 7.9321 - acc: 0.5079 - val_loss: 8.7701 - val_acc: 0.4559
Epoch 23/50
Epoch 00022: val_loss did not improve
47s - loss: 8.0281 - acc: 0.5019 - val_loss: 7.8220 - val_acc: 0.5147
Epoch 24/50
Epoch 00023: val_loss did not improve
47s - loss: 8.0753 - acc: 0.4990 - val_loss: 7.5850 - val_acc: 0.5294
Epoch 25/50
Epoch 00024: val_loss did not improve
48s - loss: 7.8964 - acc: 0.5101 - val_loss: 7.8912 - val_acc: 0.5104
Epoch 26/50
Epoch 00025: val_loss did not improve
48s - loss: 8.2557 - acc: 0.4878 - val_loss: 8.2961 - val_acc: 0.4853
Epoch 27/50
Epoch 00026: val_loss did not improve
47s - loss: 7.8477 - acc: 0.5131 - val_loss: 7.8220 - val_acc: 0.5147
Epoch 28/50
Epoch 00027: val_loss did not improve
48s - loss: 8.2542 - acc: 0.4879 - val_loss: 8.0590 - val_acc: 0.5000
Epoch 29/50
Epoch 00028: val_loss did not improve
48s - loss: 8.0931 - acc: 0.4979 - val_loss: 7.3480 - val_acc: 0.5441
Epoch 30/50
Epoch 00029: val_loss did not improve
48s - loss: 7.9924 - acc: 0.5041 - val_loss: 7.8220 - val_acc: 0.5147
Epoch 31/50
Epoch 00030: val_loss did not improve
47s - loss: 8.0900 - acc: 0.4981 - val_loss: 7.3480 - val_acc: 0.5441
Epoch 32/50
Epoch 00031: val_loss improved from 7.34796 to 7.11092, saving model to saved_models/aug_model.human.weights.best.hdf5
49s - loss: 8.0118 - acc: 0.5029 - val_loss: 7.1109 - val_acc: 0.5588
Epoch 33/50
Epoch 00032: val_loss did not improve
48s - loss: 8.0915 - acc: 0.4980 - val_loss: 7.8220 - val_acc: 0.5147
Epoch 34/50
Epoch 00033: val_loss did not improve
48s - loss: 7.9158 - acc: 0.5089 - val_loss: 7.5850 - val_acc: 0.5294
Epoch 35/50
Epoch 00034: val_loss did not improve
48s - loss: 8.0916 - acc: 0.4980 - val_loss: 8.2961 - val_acc: 0.4853
Epoch 36/50
Epoch 00035: val_loss improved from 7.11092 to 6.87389, saving model to saved_models/aug_model.human.weights.best.hdf5
49s - loss: 7.9793 - acc: 0.5049 - val_loss: 6.8739 - val_acc: 0.5735
Epoch 37/50
Epoch 00036: val_loss did not improve
47s - loss: 8.1744 - acc: 0.4928 - val_loss: 8.2961 - val_acc: 0.4853
Epoch 38/50
Epoch 00037: val_loss did not improve
47s - loss: 7.8314 - acc: 0.5141 - val_loss: 8.2961 - val_acc: 0.4853
Epoch 39/50
Epoch 00038: val_loss did not improve
47s - loss: 8.2007 - acc: 0.4912 - val_loss: 8.2961 - val_acc: 0.4853
Epoch 40/50
Epoch 00039: val_loss did not improve
51s - loss: 8.1272 - acc: 0.4958 - val_loss: 7.8220 - val_acc: 0.5147
Epoch 41/50
Epoch 00040: val_loss did not improve
52s - loss: 8.1063 - acc: 0.4971 - val_loss: 7.5850 - val_acc: 0.5294
Epoch 42/50
Epoch 00041: val_loss did not improve
51s - loss: 8.1551 - acc: 0.4940 - val_loss: 7.5850 - val_acc: 0.5294
Epoch 43/50
Epoch 00042: val_loss did not improve
49s - loss: 7.9111 - acc: 0.5092 - val_loss: 8.0590 - val_acc: 0.5000
Epoch 44/50
Epoch 00043: val_loss did not improve
48s - loss: 8.0622 - acc: 0.4998 - val_loss: 8.0590 - val_acc: 0.5000
Epoch 45/50
Epoch 00044: val_loss did not improve
47s - loss: 8.0281 - acc: 0.5019 - val_loss: 8.2961 - val_acc: 0.4853
Epoch 46/50
Epoch 00045: val_loss did not improve
48s - loss: 8.0265 - acc: 0.5020 - val_loss: 7.5850 - val_acc: 0.5294
Epoch 47/50
Epoch 00046: val_loss improved from 6.87389 to 6.16280, saving model to saved_models/aug_model.human.weights.best.hdf5
49s - loss: 7.8802 - acc: 0.5111 - val_loss: 6.1628 - val_acc: 0.6176
Epoch 48/50
Epoch 00047: val_loss did not improve
47s - loss: 8.3649 - acc: 0.4810 - val_loss: 8.0590 - val_acc: 0.5000
Epoch 49/50
Epoch 00048: val_loss did not improve
47s - loss: 7.7160 - acc: 0.5213 - val_loss: 8.0590 - val_acc: 0.5000
Epoch 50/50
Epoch 00049: val_loss improved from 6.16280 to 5.92577, saving model to saved_models/aug_model.human.weights.best.hdf5
49s - loss: 8.2883 - acc: 0.4858 - val_loss: 5.9258 - val_acc: 0.6324
Out[16]:
<keras.callbacks.History at 0x7f224c3673c8>
Epoch 1/100
Epoch 00000: val_loss improved from inf to 7.72325, saving model to aug_model.human.weights.best.hdf5
53s - loss: 8.1648 - acc: 0.4909 - val_loss: 7.7233 - val_acc: 0.5208
Epoch 2/100
Epoch 00001: val_loss did not improve
47s - loss: 8.0281 - acc: 0.5019 - val_loss: 8.5331 - val_acc: 0.4706
Epoch 3/100
Epoch 00002: val_loss improved from 7.72325 to 7.58499, saving model to aug_model.human.weights.best.hdf5
48s - loss: 7.9452 - acc: 0.5071 - val_loss: 7.5850 - val_acc: 0.5294
Epoch 4/100
Epoch 00003: val_loss did not improve
47s - loss: 8.0590 - acc: 0.5000 - val_loss: 7.5850 - val_acc: 0.5294
Epoch 5/100
Epoch 00004: val_loss did not improve
47s - loss: 8.1210 - acc: 0.4962 - val_loss: 8.0590 - val_acc: 0.5000
Epoch 6/100
Epoch 00005: val_loss did not improve
46s - loss: 8.0769 - acc: 0.4989 - val_loss: 8.0590 - val_acc: 0.5000
Epoch 7/100
Epoch 00006: val_loss did not improve
47s - loss: 7.9940 - acc: 0.5040 - val_loss: 8.0590 - val_acc: 0.5000
Epoch 8/100
Epoch 00007: val_loss improved from 7.58499 to 6.87389, saving model to aug_model.human.weights.best.hdf5
50s - loss: 8.1551 - acc: 0.4940 - val_loss: 6.8739 - val_acc: 0.5735
Epoch 9/100
Epoch 00008: val_loss did not improve
46s - loss: 8.1094 - acc: 0.4969 - val_loss: 8.2961 - val_acc: 0.4853
Epoch 10/100
Epoch 00009: val_loss did not improve
47s - loss: 7.9305 - acc: 0.5080 - val_loss: 7.8912 - val_acc: 0.5104
Epoch 11/100
Epoch 00010: val_loss did not improve
47s - loss: 8.2557 - acc: 0.4878 - val_loss: 9.7183 - val_acc: 0.3971
Epoch 12/100
Epoch 00011: val_loss did not improve
47s - loss: 7.9274 - acc: 0.5082 - val_loss: 8.0590 - val_acc: 0.5000
Epoch 13/100
Epoch 00012: val_loss did not improve
47s - loss: 7.9599 - acc: 0.5061 - val_loss: 8.2961 - val_acc: 0.4853
Epoch 14/100
Epoch 00013: val_loss did not improve
47s - loss: 8.3030 - acc: 0.4849 - val_loss: 7.5850 - val_acc: 0.5294
Epoch 15/100
Epoch 00014: val_loss did not improve
47s - loss: 7.9777 - acc: 0.5050 - val_loss: 8.0590 - val_acc: 0.5000
Epoch 16/100
Epoch 00015: val_loss did not improve
47s - loss: 8.0281 - acc: 0.5019 - val_loss: 8.5331 - val_acc: 0.4706
Epoch 17/100
Epoch 00016: val_loss did not improve
47s - loss: 7.8639 - acc: 0.5121 - val_loss: 6.8739 - val_acc: 0.5735
Epoch 18/100
Epoch 00017: val_loss did not improve
47s - loss: 8.3161 - acc: 0.4841 - val_loss: 7.1109 - val_acc: 0.5588
Epoch 19/100
Epoch 00018: val_loss did not improve
47s - loss: 8.0281 - acc: 0.5019 - val_loss: 7.3480 - val_acc: 0.5441
Epoch 20/100
Epoch 00019: val_loss did not improve
47s - loss: 7.9615 - acc: 0.5061 - val_loss: 8.2961 - val_acc: 0.4853
Epoch 21/100
Epoch 00020: val_loss did not improve
47s - loss: 7.8933 - acc: 0.5103 - val_loss: 6.8739 - val_acc: 0.5735
Epoch 22/100
Epoch 00021: val_loss did not improve
48s - loss: 8.5290 - acc: 0.4708 - val_loss: 7.5850 - val_acc: 0.5294
Epoch 23/100
Epoch 00022: val_loss did not improve
52s - loss: 7.8833 - acc: 0.5109 - val_loss: 7.8220 - val_acc: 0.5147
Epoch 24/100
Epoch 00023: val_loss did not improve
48s - loss: 7.9483 - acc: 0.5069 - val_loss: 9.2442 - val_acc: 0.4265
Epoch 25/100
Epoch 00024: val_loss did not improve
49s - loss: 8.0590 - acc: 0.5000 - val_loss: 7.8220 - val_acc: 0.5147
Epoch 26/100
Epoch 00025: val_loss did not improve
50s - loss: 8.0087 - acc: 0.5031 - val_loss: 7.3480 - val_acc: 0.5441
Epoch 27/100
Epoch 00026: val_loss did not improve
52s - loss: 7.9452 - acc: 0.5071 - val_loss: 8.2961 - val_acc: 0.4853
Epoch 28/100
Epoch 00027: val_loss did not improve
50s - loss: 8.2410 - acc: 0.4887 - val_loss: 8.3948 - val_acc: 0.4792
Epoch 29/100
Epoch 00028: val_loss did not improve
50s - loss: 8.0869 - acc: 0.4983 - val_loss: 8.0590 - val_acc: 0.5000
Epoch 30/100
Epoch 00029: val_loss did not improve
50s - loss: 7.9924 - acc: 0.5041 - val_loss: 9.0072 - val_acc: 0.4412
Epoch 31/100
Epoch 00030: val_loss did not improve
50s - loss: 8.0134 - acc: 0.5028 - val_loss: 7.8220 - val_acc: 0.5147
Epoch 32/100
Epoch 00031: val_loss did not improve
50s - loss: 8.1860 - acc: 0.4921 - val_loss: 7.3480 - val_acc: 0.5441
Epoch 33/100
Epoch 00032: val_loss did not improve
51s - loss: 8.0266 - acc: 0.5020 - val_loss: 8.7701 - val_acc: 0.4559
Epoch 34/100
Epoch 00033: val_loss did not improve
50s - loss: 8.2495 - acc: 0.4882 - val_loss: 8.0590 - val_acc: 0.5000
Epoch 35/100
Epoch 00034: val_loss did not improve
50s - loss: 7.9956 - acc: 0.5039 - val_loss: 9.2442 - val_acc: 0.4265
Epoch 36/100
Epoch 00035: val_loss did not improve
50s - loss: 8.1713 - acc: 0.4930 - val_loss: 8.7701 - val_acc: 0.4559
Epoch 37/100
Epoch 00036: val_loss did not improve
52s - loss: 8.1047 - acc: 0.4972 - val_loss: 7.5850 - val_acc: 0.5294
Epoch 38/100
Epoch 00037: val_loss did not improve
51s - loss: 8.0606 - acc: 0.4999 - val_loss: 7.5850 - val_acc: 0.5294
Epoch 39/100
Epoch 00038: val_loss did not improve
50s - loss: 8.1388 - acc: 0.4951 - val_loss: 8.7701 - val_acc: 0.4559
Epoch 40/100
Epoch 00039: val_loss did not improve
50s - loss: 7.8020 - acc: 0.5159 - val_loss: 7.8220 - val_acc: 0.5147
Epoch 41/100
Epoch 00040: val_loss did not improve
51s - loss: 8.2395 - acc: 0.4888 - val_loss: 9.0072 - val_acc: 0.4412
Epoch 42/100
Epoch 00041: val_loss did not improve
51s - loss: 7.9793 - acc: 0.5049 - val_loss: 8.2961 - val_acc: 0.4853
Epoch 43/100
Epoch 00042: val_loss did not improve
51s - loss: 8.0265 - acc: 0.5020 - val_loss: 8.2961 - val_acc: 0.4853
Epoch 44/100
Epoch 00043: val_loss improved from 6.87389 to 6.39983, saving model to aug_model.human.weights.best.hdf5
52s - loss: 7.8345 - acc: 0.5139 - val_loss: 6.3998 - val_acc: 0.6029
Epoch 45/100
Epoch 00044: val_loss did not improve
51s - loss: 8.0916 - acc: 0.4980 - val_loss: 7.3480 - val_acc: 0.5441
Epoch 46/100
Epoch 00045: val_loss did not improve
51s - loss: 8.0118 - acc: 0.5029 - val_loss: 7.8220 - val_acc: 0.5147
Epoch 47/100
Epoch 00046: val_loss did not improve
51s - loss: 8.1566 - acc: 0.4939 - val_loss: 8.7701 - val_acc: 0.4559
Epoch 48/100
Epoch 00047: val_loss did not improve
51s - loss: 8.2511 - acc: 0.4881 - val_loss: 8.5331 - val_acc: 0.4706
Epoch 49/100
Epoch 00048: val_loss did not improve
51s - loss: 7.8020 - acc: 0.5159 - val_loss: 8.5331 - val_acc: 0.4706
Epoch 50/100
Epoch 00049: val_loss did not improve
51s - loss: 8.0622 - acc: 0.4998 - val_loss: 7.5850 - val_acc: 0.5294
Epoch 51/100
Epoch 00050: val_loss did not improve
51s - loss: 7.9321 - acc: 0.5079 - val_loss: 7.5850 - val_acc: 0.5294
Epoch 52/100
Epoch 00051: val_loss did not improve
51s - loss: 8.0265 - acc: 0.5020 - val_loss: 8.0590 - val_acc: 0.5000
Epoch 53/100
Epoch 00052: val_loss did not improve
51s - loss: 8.1063 - acc: 0.4971 - val_loss: 8.2961 - val_acc: 0.4853
Epoch 54/100
Epoch 00053: val_loss did not improve
51s - loss: 7.7501 - acc: 0.5192 - val_loss: 8.2961 - val_acc: 0.4853
Epoch 55/100
Epoch 00054: val_loss did not improve
51s - loss: 8.2689 - acc: 0.4870 - val_loss: 8.7701 - val_acc: 0.4559
Epoch 56/100
Epoch 00055: val_loss did not improve
52s - loss: 8.3014 - acc: 0.4850 - val_loss: 8.7701 - val_acc: 0.4559
Epoch 57/100
Epoch 00056: val_loss did not improve
58s - loss: 7.9599 - acc: 0.5061 - val_loss: 8.5331 - val_acc: 0.4706
Epoch 58/100
Epoch 00057: val_loss did not improve
54s - loss: 8.0281 - acc: 0.5019 - val_loss: 7.8220 - val_acc: 0.5147
Epoch 59/100
Epoch 00058: val_loss did not improve
54s - loss: 7.8980 - acc: 0.5100 - val_loss: 7.5850 - val_acc: 0.5294
Epoch 60/100
Epoch 00059: val_loss did not improve
54s - loss: 8.4299 - acc: 0.4770 - val_loss: 7.3480 - val_acc: 0.5441
Epoch 61/100
Epoch 00060: val_loss did not improve
57s - loss: 7.9599 - acc: 0.5061 - val_loss: 8.2961 - val_acc: 0.4853
Epoch 62/100
Epoch 00061: val_loss did not improve
55s - loss: 8.0753 - acc: 0.4990 - val_loss: 6.8739 - val_acc: 0.5735
Epoch 63/100
Epoch 00062: val_loss did not improve
51s - loss: 8.0265 - acc: 0.5020 - val_loss: 7.8220 - val_acc: 0.5147
Epoch 64/100
Epoch 00063: val_loss did not improve
53s - loss: 8.2348 - acc: 0.4891 - val_loss: 8.0590 - val_acc: 0.5000
Epoch 65/100
Epoch 00064: val_loss did not improve
54s - loss: 8.0753 - acc: 0.4990 - val_loss: 8.0590 - val_acc: 0.5000
Epoch 66/100
Epoch 00065: val_loss did not improve
53s - loss: 7.9793 - acc: 0.5049 - val_loss: 7.8220 - val_acc: 0.5147
Epoch 67/100
Epoch 00066: val_loss did not improve
55s - loss: 8.2201 - acc: 0.4900 - val_loss: 7.5850 - val_acc: 0.5294
Epoch 68/100
Epoch 00067: val_loss did not improve
51s - loss: 8.0590 - acc: 0.5000 - val_loss: 8.0590 - val_acc: 0.5000
Epoch 69/100
Epoch 00068: val_loss did not improve
54s - loss: 8.0412 - acc: 0.5011 - val_loss: 8.7701 - val_acc: 0.4559
Epoch 70/100
Epoch 00069: val_loss did not improve
54s - loss: 7.8996 - acc: 0.5099 - val_loss: 7.8220 - val_acc: 0.5147
Epoch 71/100
Epoch 00070: val_loss did not improve
56s - loss: 8.0606 - acc: 0.4999 - val_loss: 7.1109 - val_acc: 0.5588
Epoch 72/100
Epoch 00071: val_loss did not improve
57s - loss: 8.0737 - acc: 0.4991 - val_loss: 7.8220 - val_acc: 0.5147
Epoch 73/100
Epoch 00072: val_loss did not improve
53s - loss: 8.0769 - acc: 0.4989 - val_loss: 8.0590 - val_acc: 0.5000
Epoch 74/100
Epoch 00073: val_loss did not improve
57s - loss: 8.1078 - acc: 0.4970 - val_loss: 6.8739 - val_acc: 0.5735
Epoch 75/100
Epoch 00074: val_loss did not improve
54s - loss: 8.0753 - acc: 0.4990 - val_loss: 8.7701 - val_acc: 0.4559
Epoch 76/100
Epoch 00075: val_loss did not improve
56s - loss: 8.0296 - acc: 0.5018 - val_loss: 7.5850 - val_acc: 0.5294
Epoch 77/100
Epoch 00076: val_loss did not improve
56s - loss: 8.0250 - acc: 0.5021 - val_loss: 8.2961 - val_acc: 0.4853
Epoch 78/100
Epoch 00077: val_loss did not improve
58s - loss: 7.9158 - acc: 0.5089 - val_loss: 9.0072 - val_acc: 0.4412
Epoch 79/100
Epoch 00078: val_loss did not improve
57s - loss: 8.2689 - acc: 0.4870 - val_loss: 8.5331 - val_acc: 0.4706
Epoch 80/100
Epoch 00079: val_loss did not improve
55s - loss: 7.7013 - acc: 0.5222 - val_loss: 7.8220 - val_acc: 0.5147
Epoch 81/100
Epoch 00080: val_loss did not improve
56s - loss: 8.3649 - acc: 0.4810 - val_loss: 8.2961 - val_acc: 0.4853
Epoch 82/100
Epoch 00081: val_loss did not improve
57s - loss: 7.8980 - acc: 0.5100 - val_loss: 6.8739 - val_acc: 0.5735
Epoch 83/100
Epoch 00082: val_loss did not improve
57s - loss: 7.6866 - acc: 0.5231 - val_loss: 7.3480 - val_acc: 0.5441
Epoch 84/100
Epoch 00083: val_loss did not improve
56s - loss: 8.3502 - acc: 0.4819 - val_loss: 9.0072 - val_acc: 0.4412
Epoch 85/100
Epoch 00084: val_loss did not improve
58s - loss: 8.1450 - acc: 0.4947 - val_loss: 8.2961 - val_acc: 0.4853
Epoch 86/100
Epoch 00085: val_loss did not improve
55s - loss: 7.7354 - acc: 0.5201 - val_loss: 9.0072 - val_acc: 0.4412
Epoch 87/100
Epoch 00086: val_loss did not improve
52s - loss: 8.2217 - acc: 0.4899 - val_loss: 8.7701 - val_acc: 0.4559
Epoch 88/100
Epoch 00087: val_loss did not improve
55s - loss: 8.2867 - acc: 0.4859 - val_loss: 9.0072 - val_acc: 0.4412
Epoch 89/100
Epoch 00088: val_loss did not improve
56s - loss: 7.9909 - acc: 0.5042 - val_loss: 8.0590 - val_acc: 0.5000
Epoch 90/100
Epoch 00089: val_loss did not improve
54s - loss: 8.0916 - acc: 0.4980 - val_loss: 8.0590 - val_acc: 0.5000
Epoch 91/100
Epoch 00090: val_loss did not improve
52s - loss: 7.8639 - acc: 0.5121 - val_loss: 8.0590 - val_acc: 0.5000
Epoch 92/100
Epoch 00091: val_loss did not improve
54s - loss: 8.1241 - acc: 0.4960 - val_loss: 8.2961 - val_acc: 0.4853
Epoch 93/100
Epoch 00092: val_loss did not improve
54s - loss: 8.2836 - acc: 0.4861 - val_loss: 8.5331 - val_acc: 0.4706
Epoch 94/100
Epoch 00093: val_loss did not improve
54s - loss: 7.9762 - acc: 0.5051 - val_loss: 8.0590 - val_acc: 0.5000
Epoch 95/100
Epoch 00094: val_loss did not improve
57s - loss: 8.1404 - acc: 0.4950 - val_loss: 8.5331 - val_acc: 0.4706
Epoch 96/100
Epoch 00095: val_loss did not improve
53s - loss: 8.0737 - acc: 0.4991 - val_loss: 7.3480 - val_acc: 0.5441
Epoch 97/100
Epoch 00096: val_loss did not improve
54s - loss: 8.0428 - acc: 0.5010 - val_loss: 8.5331 - val_acc: 0.4706
Epoch 98/100
Epoch 00097: val_loss did not improve
54s - loss: 8.0412 - acc: 0.5011 - val_loss: 8.2961 - val_acc: 0.4853
Epoch 99/100
Epoch 00098: val_loss did not improve
53s - loss: 8.1713 - acc: 0.4930 - val_loss: 8.5331 - val_acc: 0.4706
Epoch 100/100
Epoch 00099: val_loss did not improve
57s - loss: 8.1257 - acc: 0.4959 - val_loss: 8.2961 - val_acc: 0.4853
Out[29]:

In [17]:
# load the weights that yielded the best validation accuracy
model.load_weights('saved_models/aug_model.human.weights.best.hdf5')

# evaluate and print test accuracy
score = model.evaluate(x_test, y_test, verbose=0)
print('\n', 'Test accuracy:', score[1])
 Test accuracy: 0.5

Results:

The MLP seems to take up a really large number of parameters, but quickly settles on to a val_loss value (8.0590) that does not improve over subsequent epochs beyond the 2nd epoch. This I tried with various combinations of 3-deep networks. None of the trials gives loss below 8.0590, or test accuracy better than 0.5, which is really the same as mindlessly predicting all the images to be of one class or the other. The accuracy does not change either with different shuffling of images, or by increasing the training data from 200 to 500 images.

Results of the Conv2D, even thought the loss is much lesser at 6.39983, don't look much different, with the accuracy still being 0.5 - effectively random.


Step 2: Detect Dogs

In this section, we use a pre-trained ResNet-50 model to detect dogs in images. Our first line of code downloads the ResNet-50 model, along with weights that have been trained on ImageNet, a very large, very popular dataset used for image classification and other vision tasks. ImageNet contains over 10 million URLs, each linking to an image containing an object from one of 1000 categories. Given an image, this pre-trained ResNet-50 model returns a prediction (derived from the available categories in ImageNet) for the object that is contained in the image.

In [18]:
from keras.applications.resnet50 import ResNet50

# define ResNet50 model
ResNet50_model = ResNet50(weights='imagenet')

Pre-process the Data

When using TensorFlow as backend, Keras CNNs require a 4D array (which we'll also refer to as a 4D tensor) as input, with shape

$$ (\text{nb_samples}, \text{rows}, \text{columns}, \text{channels}), $$

where nb_samples corresponds to the total number of images (or samples), and rows, columns, and channels correspond to the number of rows, columns, and channels for each image, respectively.

The path_to_tensor function below takes a string-valued file path to a color image as input and returns a 4D tensor suitable for supplying to a Keras CNN. The function first loads the image and resizes it to a square image that is $224 \times 224$ pixels. Next, the image is converted to an array, which is then resized to a 4D tensor. In this case, since we are working with color images, each image has three channels. Likewise, since we are processing a single image (or sample), the returned tensor will always have shape

$$ (1, 224, 224, 3). $$

The paths_to_tensor function takes a numpy array of string-valued image paths as input and returns a 4D tensor with shape

$$ (\text{nb_samples}, 224, 224, 3). $$

Here, nb_samples is the number of samples, or number of images, in the supplied array of image paths. It is best to think of nb_samples as the number of 3D tensors (where each 3D tensor corresponds to a different image) in your dataset!

In [19]:
from keras.preprocessing import image                  
from tqdm import tqdm

def path_to_tensor(img_path):
    # loads RGB image as PIL.Image.Image type
    img = image.load_img(img_path, target_size=(224, 224))
    # convert PIL.Image.Image type to 3D tensor with shape (224, 224, 3)
    x = image.img_to_array(img)
    # convert 3D tensor to 4D tensor with shape (1, 224, 224, 3) and return 4D tensor
    return np.expand_dims(x, axis=0)

def paths_to_tensor(img_paths):
    list_of_tensors = [path_to_tensor(img_path) for img_path in tqdm(img_paths)]
    return np.vstack(list_of_tensors)

Making Predictions with ResNet-50

Getting the 4D tensor ready for ResNet-50, and for any other pre-trained model in Keras, requires some additional processing. First, the RGB image is converted to BGR by reordering the channels. All pre-trained models have the additional normalization step that the mean pixel (expressed in RGB as $[103.939, 116.779, 123.68]$ and calculated from all pixels in all images in ImageNet) must be subtracted from every pixel in each image. This is implemented in the imported function preprocess_input. If you're curious, you can check the code for preprocess_input here.

Now that we have a way to format our image for supplying to ResNet-50, we are now ready to use the model to extract the predictions. This is accomplished with the predict method, which returns an array whose $i$-th entry is the model's predicted probability that the image belongs to the $i$-th ImageNet category. This is implemented in the ResNet50_predict_labels function below.

By taking the argmax of the predicted probability vector, we obtain an integer corresponding to the model's predicted object class, which we can identify with an object category through the use of this dictionary.

In [20]:
from keras.applications.resnet50 import preprocess_input, decode_predictions

def ResNet50_predict_labels(img_path):
    # returns prediction vector for image located at img_path
    img = preprocess_input(path_to_tensor(img_path))
    return np.argmax(ResNet50_model.predict(img))

Write a Dog Detector

While looking at the dictionary, you will notice that the categories corresponding to dogs appear in an uninterrupted sequence and correspond to dictionary keys 151-268, inclusive, to include all categories from 'Chihuahua' to 'Mexican hairless'. Thus, in order to check to see if an image is predicted to contain a dog by the pre-trained ResNet-50 model, we need only check if the ResNet50_predict_labels function above returns a value between 151 and 268 (inclusive).

We use these ideas to complete the dog_detector function below, which returns True if a dog is detected in an image (and False if not).

In [21]:
### returns "True" if a dog is detected in the image stored at img_path
def dog_detector(img_path):
    prediction = ResNet50_predict_labels(img_path)
    #print if any human resembles a canine
    print(img_path) if ((prediction <= 268) & (prediction >= 151)) and 'lfw' in img_path else None
    return ((prediction <= 268) & (prediction >= 151)) 

(IMPLEMENTATION) Assess the Dog Detector

Question 3: Use the code cell below to test the performance of your dog_detector function.

  • What percentage of the images in human_files_short have a detected dog?
  • What percentage of the images in dog_files_short have a detected dog?

Answer:

  • What percentage of the images in human_files_short have a detected dog?
     Only 1% human faces were incorrectly reccognized as dog-face
     Different runs with different random seed turn up 4 issues... 
  • What percentage of the images in dog_files_short have a detected dog?
      100% of the dog faces were detected correctly by the ResNet-50 Dog Detector.
In [22]:
### TODO: Test the performance of the dog_detector function
### on the images in human_files_short and dog_files_short.

hres_det = list(map(dog_detector, human_files_short))
dres_det = list(map(dog_detector, dog_files_short))
lfw/Andrew_Fastow/Andrew_Fastow_0001.jpg
lfw/George_W_Bush/George_W_Bush_0187.jpg
lfw/Doris_Schroeder/Doris_Schroeder_0001.jpg
lfw/Roy_Williams/Roy_Williams_0001.jpg
In [23]:
print ('{}\n{}'.format(sum(hres_det),sum(dres_det)))
4
100

Step 3: Create a CNN to Classify Dog Breeds (from Scratch)

Now that we have functions for detecting humans and dogs in images, we need a way to predict breed from images. In this step, you will create a CNN that classifies dog breeds. You must create your CNN from scratch (so, you can't use transfer learning yet!), and you must attain a test accuracy of at least 1%. In Step 5 of this notebook, you will have the opportunity to use transfer learning to create a CNN that attains greatly improved accuracy.

Be careful with adding too many trainable layers! More parameters means longer training, which means you are more likely to need a GPU to accelerate the training process. Thankfully, Keras provides a handy estimate of the time that each epoch is likely to take; you can extrapolate this estimate to figure out how long it will take for your algorithm to train.

We mention that the task of assigning breed to dogs from images is considered exceptionally challenging. To see why, consider that even a human would have great difficulty in distinguishing between a Brittany and a Welsh Springer Spaniel.

Brittany Welsh Springer Spaniel

It is not difficult to find other dog breed pairs with minimal inter-class variation (for instance, Curly-Coated Retrievers and American Water Spaniels).

Curly-Coated Retriever American Water Spaniel

Likewise, recall that labradors come in yellow, chocolate, and black. Your vision-based algorithm will have to conquer this high intra-class variation to determine how to classify all of these different shades as the same breed.

Yellow Labrador Chocolate Labrador Black Labrador

We also mention that random chance presents an exceptionally low bar: setting aside the fact that the classes are slightly imabalanced, a random guess will provide a correct answer roughly 1 in 133 times, which corresponds to an accuracy of less than 1%.

Remember that the practice is far ahead of the theory in deep learning. Experiment with many different architectures, and trust your intuition. And, of course, have fun!

Pre-process the Data

We rescale the images by dividing every pixel in every image by 255.

In [24]:
from PIL import ImageFile                            
ImageFile.LOAD_TRUNCATED_IMAGES = True                 

# pre-process the data for Keras
train_tensors = paths_to_tensor(train_files).astype('float32')/255
valid_tensors = paths_to_tensor(valid_files).astype('float32')/255
test_tensors = paths_to_tensor(test_files).astype('float32')/255
100%|██████████| 6680/6680 [00:47<00:00, 141.42it/s]
100%|██████████| 835/835 [00:05<00:00, 149.81it/s]
100%|██████████| 836/836 [00:05<00:00, 161.56it/s]

(IMPLEMENTATION) Model Architecture

Create a CNN to classify dog breed. At the end of your code cell block, summarize the layers of your model by executing the line:

    model.summary()

We have imported some Python modules to get you started, but feel free to import as many modules as you need. If you end up getting stuck, here's a hint that specifies a model that trains relatively fast on CPU and attains >1% test accuracy in 5 epochs:

Sample CNN

Question 4: Outline the steps you took to get to your final CNN architecture and your reasoning at each step. If you chose to use the hinted architecture above, describe why you think that CNN architecture should work well for the image classification task.

Answer:

As mentioned in the videos, the AlexNet was the first one to introduce the Conv-Pool-Dropout architecture with ReLU activation. While computationally intensive this is known to give good results in the image recognition arena. It would be instructional to try out the VGG Architecture along with this one as well.

I use the same exact architecture as suggested above, using 'valid' padding. This is different from the one that I used earlier in the Conv2D example. While keeping the layer sizes as given in the model summary above, I've added 3 Dropout layers. To begin with, I started with a value of 0.2 for the drop-rate. The results though seem to be pretty nice, giving an accuracy of 4.18%, much better than required 1%.

In [25]:
from keras.layers import Conv2D, MaxPooling2D, GlobalAveragePooling2D
from keras.layers import Dropout, Flatten, Dense
from keras.models import Sequential

model = Sequential()

### TODO: Define your architecture.
# Start off with the one in the other miniproject (or the same used above!)
# Then modify according to the hint given...

model = Sequential()
model.add(Conv2D(filters=16, kernel_size=2, padding='valid', activation='relu', 
                        input_shape=(224, 224, 3)))
model.add(MaxPooling2D(pool_size=2))
model.add(Dropout(0.2))

model.add(Conv2D(filters=32, kernel_size=2, padding='valid', activation='relu'))
model.add(MaxPooling2D(pool_size=2))
model.add(Dropout(0.2))

model.add(Conv2D(filters=64, kernel_size=2, padding='valid', activation='relu'))
model.add(MaxPooling2D(pool_size=2))
model.add(Dropout(0.2))

model.add(GlobalAveragePooling2D())
model.add(Dense(133, activation='softmax'))

model.summary()
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
conv2d_4 (Conv2D)            (None, 223, 223, 16)      208       
_________________________________________________________________
max_pooling2d_5 (MaxPooling2 (None, 111, 111, 16)      0         
_________________________________________________________________
dropout_5 (Dropout)          (None, 111, 111, 16)      0         
_________________________________________________________________
conv2d_5 (Conv2D)            (None, 110, 110, 32)      2080      
_________________________________________________________________
max_pooling2d_6 (MaxPooling2 (None, 55, 55, 32)        0         
_________________________________________________________________
dropout_6 (Dropout)          (None, 55, 55, 32)        0         
_________________________________________________________________
conv2d_6 (Conv2D)            (None, 54, 54, 64)        8256      
_________________________________________________________________
max_pooling2d_7 (MaxPooling2 (None, 27, 27, 64)        0         
_________________________________________________________________
dropout_7 (Dropout)          (None, 27, 27, 64)        0         
_________________________________________________________________
global_average_pooling2d_1 ( (None, 64)                0         
_________________________________________________________________
dense_7 (Dense)              (None, 133)               8645      
=================================================================
Total params: 19,189.0
Trainable params: 19,189.0
Non-trainable params: 0.0
_________________________________________________________________

Compile the Model

In [26]:
model.compile(optimizer='rmsprop', loss='categorical_crossentropy', metrics=['accuracy'])

(IMPLEMENTATION) Train the Model

Train your model in the code cell below. Use model checkpointing to save the model that attains the best validation loss.

You are welcome to augment the training data, but this is not a requirement.

In [27]:
from keras.callbacks import ModelCheckpoint  

### TODO: specify the number of epochs that you would like to use to train the model.
# takes about 1:10 to run
epochs = 20

### Do NOT modify the code below this line.

checkpointer = ModelCheckpoint(filepath='saved_models/weights.best.from_scratch.hdf5', 
                               verbose=1, save_best_only=True)

model.fit(train_tensors, train_targets, 
          validation_data=(valid_tensors, valid_targets),
          epochs=epochs, batch_size=20, callbacks=[checkpointer], verbose=1)
Train on 6680 samples, validate on 835 samples
Epoch 1/20
6660/6680 [============================>.] - ETA: 0s - loss: 4.8829 - acc: 0.0105Epoch 00000: val_loss improved from inf to 4.87696, saving model to saved_models/weights.best.from_scratch.hdf5
6680/6680 [==============================] - 210s - loss: 4.8829 - acc: 0.0105 - val_loss: 4.8770 - val_acc: 0.0108
Epoch 2/20
6660/6680 [============================>.] - ETA: 0s - loss: 4.8473 - acc: 0.0131Epoch 00001: val_loss improved from 4.87696 to 4.82448, saving model to saved_models/weights.best.from_scratch.hdf5
6680/6680 [==============================] - 210s - loss: 4.8471 - acc: 0.0130 - val_loss: 4.8245 - val_acc: 0.0204
Epoch 3/20
6660/6680 [============================>.] - ETA: 0s - loss: 4.8037 - acc: 0.0159Epoch 00002: val_loss improved from 4.82448 to 4.80281, saving model to saved_models/weights.best.from_scratch.hdf5
6680/6680 [==============================] - 220s - loss: 4.8035 - acc: 0.0159 - val_loss: 4.8028 - val_acc: 0.0180
Epoch 4/20
6660/6680 [============================>.] - ETA: 0s - loss: 4.7794 - acc: 0.0185Epoch 00003: val_loss improved from 4.80281 to 4.78950, saving model to saved_models/weights.best.from_scratch.hdf5
6680/6680 [==============================] - 206s - loss: 4.7799 - acc: 0.0184 - val_loss: 4.7895 - val_acc: 0.0180
Epoch 5/20
6660/6680 [============================>.] - ETA: 0s - loss: 4.7601 - acc: 0.0212Epoch 00004: val_loss improved from 4.78950 to 4.77705, saving model to saved_models/weights.best.from_scratch.hdf5
6680/6680 [==============================] - 204s - loss: 4.7594 - acc: 0.0214 - val_loss: 4.7770 - val_acc: 0.0156
Epoch 6/20
6660/6680 [============================>.] - ETA: 0s - loss: 4.7382 - acc: 0.0222Epoch 00005: val_loss improved from 4.77705 to 4.75554, saving model to saved_models/weights.best.from_scratch.hdf5
6680/6680 [==============================] - 209s - loss: 4.7383 - acc: 0.0222 - val_loss: 4.7555 - val_acc: 0.0168
Epoch 7/20
6660/6680 [============================>.] - ETA: 0s - loss: 4.7146 - acc: 0.0254Epoch 00006: val_loss improved from 4.75554 to 4.74786, saving model to saved_models/weights.best.from_scratch.hdf5
6680/6680 [==============================] - 210s - loss: 4.7143 - acc: 0.0253 - val_loss: 4.7479 - val_acc: 0.0192
Epoch 8/20
6660/6680 [============================>.] - ETA: 0s - loss: 4.6932 - acc: 0.0243Epoch 00007: val_loss improved from 4.74786 to 4.72532, saving model to saved_models/weights.best.from_scratch.hdf5
6680/6680 [==============================] - 202s - loss: 4.6937 - acc: 0.0243 - val_loss: 4.7253 - val_acc: 0.0216
Epoch 9/20
6660/6680 [============================>.] - ETA: 0s - loss: 4.6743 - acc: 0.0282Epoch 00008: val_loss improved from 4.72532 to 4.70654, saving model to saved_models/weights.best.from_scratch.hdf5
6680/6680 [==============================] - 202s - loss: 4.6741 - acc: 0.0284 - val_loss: 4.7065 - val_acc: 0.0192
Epoch 10/20
6660/6680 [============================>.] - ETA: 0s - loss: 4.6565 - acc: 0.0332Epoch 00009: val_loss improved from 4.70654 to 4.70058, saving model to saved_models/weights.best.from_scratch.hdf5
6680/6680 [==============================] - 202s - loss: 4.6567 - acc: 0.0331 - val_loss: 4.7006 - val_acc: 0.0335
Epoch 11/20
6660/6680 [============================>.] - ETA: 0s - loss: 4.6425 - acc: 0.0299Epoch 00010: val_loss improved from 4.70058 to 4.68611, saving model to saved_models/weights.best.from_scratch.hdf5
6680/6680 [==============================] - 203s - loss: 4.6422 - acc: 0.0299 - val_loss: 4.6861 - val_acc: 0.0311
Epoch 12/20
6660/6680 [============================>.] - ETA: 0s - loss: 4.6252 - acc: 0.0339Epoch 00011: val_loss improved from 4.68611 to 4.67870, saving model to saved_models/weights.best.from_scratch.hdf5
6680/6680 [==============================] - 210s - loss: 4.6249 - acc: 0.0341 - val_loss: 4.6787 - val_acc: 0.0371
Epoch 13/20
6660/6680 [============================>.] - ETA: 0s - loss: 4.6052 - acc: 0.0383Epoch 00012: val_loss improved from 4.67870 to 4.66766, saving model to saved_models/weights.best.from_scratch.hdf5
6680/6680 [==============================] - 224s - loss: 4.6058 - acc: 0.0382 - val_loss: 4.6677 - val_acc: 0.0311
Epoch 14/20
6660/6680 [============================>.] - ETA: 0s - loss: 4.5910 - acc: 0.0396Epoch 00013: val_loss improved from 4.66766 to 4.66626, saving model to saved_models/weights.best.from_scratch.hdf5
6680/6680 [==============================] - 216s - loss: 4.5921 - acc: 0.0397 - val_loss: 4.6663 - val_acc: 0.0407
Epoch 15/20
6660/6680 [============================>.] - ETA: 0s - loss: 4.5703 - acc: 0.0399Epoch 00014: val_loss improved from 4.66626 to 4.63283, saving model to saved_models/weights.best.from_scratch.hdf5
6680/6680 [==============================] - 216s - loss: 4.5702 - acc: 0.0400 - val_loss: 4.6328 - val_acc: 0.0299
Epoch 16/20
6660/6680 [============================>.] - ETA: 0s - loss: 4.5566 - acc: 0.0447Epoch 00015: val_loss improved from 4.63283 to 4.62061, saving model to saved_models/weights.best.from_scratch.hdf5
6680/6680 [==============================] - 232s - loss: 4.5565 - acc: 0.0446 - val_loss: 4.6206 - val_acc: 0.0311
Epoch 17/20
6660/6680 [============================>.] - ETA: 0s - loss: 4.5326 - acc: 0.0480Epoch 00016: val_loss improved from 4.62061 to 4.61475, saving model to saved_models/weights.best.from_scratch.hdf5
6680/6680 [==============================] - 230s - loss: 4.5331 - acc: 0.0479 - val_loss: 4.6148 - val_acc: 0.0347
Epoch 18/20
6660/6680 [============================>.] - ETA: 0s - loss: 4.5158 - acc: 0.0492Epoch 00017: val_loss improved from 4.61475 to 4.60218, saving model to saved_models/weights.best.from_scratch.hdf5
6680/6680 [==============================] - 239s - loss: 4.5156 - acc: 0.0491 - val_loss: 4.6022 - val_acc: 0.0311
Epoch 19/20
6660/6680 [============================>.] - ETA: 0s - loss: 4.4964 - acc: 0.0495Epoch 00018: val_loss did not improve
6680/6680 [==============================] - 209s - loss: 4.4968 - acc: 0.0496 - val_loss: 4.6123 - val_acc: 0.0311
Epoch 20/20
6660/6680 [============================>.] - ETA: 0s - loss: 4.4850 - acc: 0.0526Epoch 00019: val_loss improved from 4.60218 to 4.57811, saving model to saved_models/weights.best.from_scratch.hdf5
6680/6680 [==============================] - 224s - loss: 4.4849 - acc: 0.0527 - val_loss: 4.5781 - val_acc: 0.0347
Out[27]:
<keras.callbacks.History at 0x7ff5b4441898>

Load the Model with the Best Validation Loss

In [28]:
model.load_weights('saved_models/weights.best.from_scratch.hdf5')

Test the Model

Try out your model on the test dataset of dog images. Ensure that your test accuracy is greater than 1%.

In [29]:
# get index of predicted dog breed for each image in test set
dog_breed_predictions = [np.argmax(model.predict(np.expand_dims(tensor, axis=0))) for tensor in test_tensors]

# report test accuracy
test_accuracy = 100*np.sum(np.array(dog_breed_predictions)==np.argmax(test_targets, axis=1))/len(dog_breed_predictions)
print('Test accuracy: %.4f%%' % test_accuracy)
Test accuracy: 4.7847%

Step 4: Use a CNN to Classify Dog Breeds

To reduce training time without sacrificing accuracy, we show you how to train a CNN using transfer learning. In the following step, you will get a chance to use transfer learning to train your own CNN.

Obtain Bottleneck Features

In [30]:
bottleneck_features = np.load('bottleneck_features/DogVGG16Data.npz')
train_VGG16 = bottleneck_features['train']
valid_VGG16 = bottleneck_features['valid']
test_VGG16 = bottleneck_features['test']

Model Architecture

The model uses the the pre-trained VGG-16 model as a fixed feature extractor, where the last convolutional output of VGG-16 is fed as input to our model. We only add a global average pooling layer and a fully connected layer, where the latter contains one node for each dog category and is equipped with a softmax.

In [31]:
VGG16_model = Sequential()
VGG16_model.add(GlobalAveragePooling2D(input_shape=train_VGG16.shape[1:]))
VGG16_model.add(Dense(133, activation='softmax'))

VGG16_model.summary()
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
global_average_pooling2d_2 ( (None, 512)               0         
_________________________________________________________________
dense_8 (Dense)              (None, 133)               68229     
=================================================================
Total params: 68,229.0
Trainable params: 68,229.0
Non-trainable params: 0.0
_________________________________________________________________

Compile the Model

In [32]:
VGG16_model.compile(loss='categorical_crossentropy', optimizer='rmsprop', metrics=['accuracy'])

Train the Model

In [33]:
from keras.callbacks import ModelCheckpoint  

checkpointer = ModelCheckpoint(filepath='saved_models/weights.best.VGG16.hdf5', 
                               verbose=1, save_best_only=True)

VGG16_model.fit(train_VGG16, train_targets, 
          validation_data=(valid_VGG16, valid_targets),
          epochs=20, batch_size=20, callbacks=[checkpointer], verbose=1)
Train on 6680 samples, validate on 835 samples
Epoch 1/20
6640/6680 [============================>.] - ETA: 0s - loss: 11.9129 - acc: 0.1325Epoch 00000: val_loss improved from inf to 10.21519, saving model to saved_models/weights.best.VGG16.hdf5
6680/6680 [==============================] - 1s - loss: 11.9067 - acc: 0.1331 - val_loss: 10.2152 - val_acc: 0.2443
Epoch 2/20
6460/6680 [============================>.] - ETA: 0s - loss: 9.5631 - acc: 0.3085Epoch 00001: val_loss improved from 10.21519 to 9.56649, saving model to saved_models/weights.best.VGG16.hdf5
6680/6680 [==============================] - 1s - loss: 9.5751 - acc: 0.3093 - val_loss: 9.5665 - val_acc: 0.3102
Epoch 3/20
6540/6680 [============================>.] - ETA: 0s - loss: 9.0809 - acc: 0.3725Epoch 00002: val_loss improved from 9.56649 to 9.30583, saving model to saved_models/weights.best.VGG16.hdf5
6680/6680 [==============================] - 1s - loss: 9.0748 - acc: 0.3723 - val_loss: 9.3058 - val_acc: 0.3365
Epoch 4/20
6540/6680 [============================>.] - ETA: 0s - loss: 8.8904 - acc: 0.4054Epoch 00003: val_loss improved from 9.30583 to 9.16529, saving model to saved_models/weights.best.VGG16.hdf5
6680/6680 [==============================] - 1s - loss: 8.8843 - acc: 0.4049 - val_loss: 9.1653 - val_acc: 0.3605
Epoch 5/20
6660/6680 [============================>.] - ETA: 0s - loss: 8.5549 - acc: 0.4218- ETA: 1s - Epoch 00004: val_loss improved from 9.16529 to 8.86390, saving model to saved_models/weights.best.VGG16.hdf5
6680/6680 [==============================] - 1s - loss: 8.5526 - acc: 0.4219 - val_loss: 8.8639 - val_acc: 0.3784
Epoch 6/20
6580/6680 [============================>.] - ETA: 0s - loss: 8.1966 - acc: 0.4523Epoch 00005: val_loss improved from 8.86390 to 8.77595, saving model to saved_models/weights.best.VGG16.hdf5
6680/6680 [==============================] - 1s - loss: 8.1999 - acc: 0.4521 - val_loss: 8.7759 - val_acc: 0.3760
Epoch 7/20
6540/6680 [============================>.] - ETA: 0s - loss: 8.0923 - acc: 0.4714Epoch 00006: val_loss improved from 8.77595 to 8.71069, saving model to saved_models/weights.best.VGG16.hdf5
6680/6680 [==============================] - 1s - loss: 8.0759 - acc: 0.4722 - val_loss: 8.7107 - val_acc: 0.3868
Epoch 8/20
6640/6680 [============================>.] - ETA: 0s - loss: 7.9688 - acc: 0.4837Epoch 00007: val_loss improved from 8.71069 to 8.46338, saving model to saved_models/weights.best.VGG16.hdf5
6680/6680 [==============================] - 1s - loss: 7.9674 - acc: 0.4837 - val_loss: 8.4634 - val_acc: 0.3976
Epoch 9/20
6540/6680 [============================>.] - ETA: 0s - loss: 7.7506 - acc: 0.4966Epoch 00008: val_loss improved from 8.46338 to 8.36460, saving model to saved_models/weights.best.VGG16.hdf5
6680/6680 [==============================] - 1s - loss: 7.7523 - acc: 0.4963 - val_loss: 8.3646 - val_acc: 0.4048
Epoch 10/20
6480/6680 [============================>.] - ETA: 0s - loss: 7.6240 - acc: 0.5040Epoch 00009: val_loss improved from 8.36460 to 8.23367, saving model to saved_models/weights.best.VGG16.hdf5
6680/6680 [==============================] - 1s - loss: 7.5898 - acc: 0.5061 - val_loss: 8.2337 - val_acc: 0.3988
Epoch 11/20
6600/6680 [============================>.] - ETA: 0s - loss: 7.4700 - acc: 0.5173Epoch 00010: val_loss improved from 8.23367 to 8.10574, saving model to saved_models/weights.best.VGG16.hdf5
6680/6680 [==============================] - 1s - loss: 7.4585 - acc: 0.5181 - val_loss: 8.1057 - val_acc: 0.4299
Epoch 12/20
6440/6680 [===========================>..] - ETA: 0s - loss: 7.3659 - acc: 0.5255Epoch 00011: val_loss improved from 8.10574 to 8.09140, saving model to saved_models/weights.best.VGG16.hdf5
6680/6680 [==============================] - 1s - loss: 7.3675 - acc: 0.5253 - val_loss: 8.0914 - val_acc: 0.4216
Epoch 13/20
6640/6680 [============================>.] - ETA: 0s - loss: 7.2744 - acc: 0.5309Epoch 00012: val_loss improved from 8.09140 to 8.06982, saving model to saved_models/weights.best.VGG16.hdf5
6680/6680 [==============================] - 1s - loss: 7.2656 - acc: 0.5314 - val_loss: 8.0698 - val_acc: 0.4180
Epoch 14/20
6560/6680 [============================>.] - ETA: 0s - loss: 7.1527 - acc: 0.5387Epoch 00013: val_loss improved from 8.06982 to 8.02316, saving model to saved_models/weights.best.VGG16.hdf5
6680/6680 [==============================] - 1s - loss: 7.1526 - acc: 0.5388 - val_loss: 8.0232 - val_acc: 0.4275
Epoch 15/20
6400/6680 [===========================>..] - ETA: 0s - loss: 7.0089 - acc: 0.5467Epoch 00014: val_loss improved from 8.02316 to 7.78861, saving model to saved_models/weights.best.VGG16.hdf5
6680/6680 [==============================] - 1s - loss: 6.9981 - acc: 0.5473 - val_loss: 7.7886 - val_acc: 0.4383
Epoch 16/20
6540/6680 [============================>.] - ETA: 0s - loss: 6.8697 - acc: 0.5598Epoch 00015: val_loss did not improve
6680/6680 [==============================] - 1s - loss: 6.8763 - acc: 0.5593 - val_loss: 7.7906 - val_acc: 0.4407
Epoch 17/20
6580/6680 [============================>.] - ETA: 0s - loss: 6.8053 - acc: 0.5647Epoch 00016: val_loss improved from 7.78861 to 7.68637, saving model to saved_models/weights.best.VGG16.hdf5
6680/6680 [==============================] - 1s - loss: 6.8297 - acc: 0.5633 - val_loss: 7.6864 - val_acc: 0.4455
Epoch 18/20
6480/6680 [============================>.] - ETA: 0s - loss: 6.7499 - acc: 0.5660Epoch 00017: val_loss did not improve
6680/6680 [==============================] - 1s - loss: 6.7712 - acc: 0.5644 - val_loss: 7.7421 - val_acc: 0.4371
Epoch 19/20
6500/6680 [============================>.] - ETA: 0s - loss: 6.5616 - acc: 0.5751Epoch 00018: val_loss improved from 7.68637 to 7.56232, saving model to saved_models/weights.best.VGG16.hdf5
6680/6680 [==============================] - 1s - loss: 6.5631 - acc: 0.5750 - val_loss: 7.5623 - val_acc: 0.4467
Epoch 20/20
6460/6680 [============================>.] - ETA: 0s - loss: 6.4629 - acc: 0.5872Epoch 00019: val_loss improved from 7.56232 to 7.52164, saving model to saved_models/weights.best.VGG16.hdf5
6680/6680 [==============================] - 1s - loss: 6.4654 - acc: 0.5868 - val_loss: 7.5216 - val_acc: 0.4467
Out[33]:
<keras.callbacks.History at 0x7ff5b41c0ac8>

Load the Model with the Best Validation Loss

In [34]:
VGG16_model.load_weights('saved_models/weights.best.VGG16.hdf5')

Test the Model

Now, we can use the CNN to test how well it identifies breed within our test dataset of dog images. We print the test accuracy below.

In [35]:
# get index of predicted dog breed for each image in test set
VGG16_predictions = [np.argmax(VGG16_model.predict(np.expand_dims(feature, axis=0))) for feature in test_VGG16]

# report test accuracy
test_accuracy = 100*np.sum(np.array(VGG16_predictions)==np.argmax(test_targets, axis=1))/len(VGG16_predictions)
print('Test accuracy: %.4f%%' % test_accuracy)
Test accuracy: 46.1722%

Predict Dog Breed with the Model

In [36]:
from extract_bottleneck_features import *

def VGG16_predict_breed(img_path):
    # extract bottleneck features
    bottleneck_feature = extract_VGG16(path_to_tensor(img_path))
    # obtain predicted vector
    predicted_vector = VGG16_model.predict(bottleneck_feature)
    # return dog breed that is predicted by the model
    return dog_names[np.argmax(predicted_vector)]
In [37]:
import random
randomlist = random.sample(range(len(test_files)), 5)

for idx in randomlist:
    print('{:3d} {:25.25} {:25.25}'.format(idx, VGG16_predict_breed(test_files[idx]), dog_names[np.argmax(test_targets[idx])]))
114 Akita                     German_shepherd_dog      
338 Kuvasz                    Kuvasz                   
293 Dogue_de_bordeaux         Dogue_de_bordeaux        
404 Kerry_blue_terrier        Kerry_blue_terrier       
124 Nova_scotia_duck_tolling_ Brittany                 

Step 5: Create a CNN to Classify Dog Breeds (using Transfer Learning)

You will now use transfer learning to create a CNN that can identify dog breed from images. Your CNN must attain at least 60% accuracy on the test set.

In Step 4, we used transfer learning to create a CNN using VGG-16 bottleneck features. In this section, you must use the bottleneck features from a different pre-trained model. To make things easier for you, we have pre-computed the features for all of the networks that are currently available in Keras:

The files are encoded as such:

Dog{network}Data.npz

where {network}, in the above filename, can be one of VGG19, Resnet50, InceptionV3, or Xception. Pick one of the above architectures, download the corresponding bottleneck features, and store the downloaded file in the bottleneck_features/ folder in the repository.

(IMPLEMENTATION) Obtain Bottleneck Features

In the code block below, extract the bottleneck features corresponding to the train, test, and validation sets by running the following:

bottleneck_features = np.load('bottleneck_features/Dog{network}Data.npz')
train_{network} = bottleneck_features['train']
valid_{network} = bottleneck_features['valid']
test_{network} = bottleneck_features['test']
In [51]:
### TODO: Obtain bottleneck features from another pre-trained CNN.
# Reason to use Resnet50 - its already downloaded, Inc=1.6GB, eXc=3.1GB not done yet.

bottleneck_features = np.load('bottleneck_features/DogResnet50Data.npz')
train_Resnet50 = bottleneck_features['train']
valid_Resnet50 = bottleneck_features['valid']
test_Resnet50 = bottleneck_features['test']

(IMPLEMENTATION) Model Architecture

Create a CNN to classify dog breed. At the end of your code cell block, summarize the layers of your model by executing the line:

    <your model's name>.summary()

Question 5: Outline the steps you took to get to your final CNN architecture and your reasoning at each step. Describe why you think the architecture is suitable for the current problem.

Answer:

We use transfer learning, i.e. pre-tuned set of parameters, to identify the features.

Since we used the similar architecture in the previous step, we'll re-use the same approach (and code!) with the Resnet50 dataset. In the Resnet50 model, the dog features are squeezed into a 2048 vector (GlobalAveragePooling2D) and then a fully-connected layer is used to obtain the predicted probabilities (133 possible breeds).

In [52]:
### TODO: Define your architecture.

Resnet50_model = Sequential()
Resnet50_model.add(GlobalAveragePooling2D(input_shape=train_Resnet50.shape[1:]))
Resnet50_model.add(Dense(133, activation='softmax'))

Resnet50_model.summary()
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
global_average_pooling2d_4 ( (None, 2048)              0         
_________________________________________________________________
dense_10 (Dense)             (None, 133)               272517    
=================================================================
Total params: 272,517.0
Trainable params: 272,517.0
Non-trainable params: 0.0
_________________________________________________________________

(IMPLEMENTATION) Compile the Model

In [53]:
### TODO: Compile the model.

Resnet50_model.compile(loss='categorical_crossentropy', optimizer='rmsprop', metrics=['accuracy'])

(IMPLEMENTATION) Train the Model

Train your model in the code cell below. Use model checkpointing to save the model that attains the best validation loss.

You are welcome to augment the training data, but this is not a requirement.

In [54]:
### TODO: Train the model.

from keras.callbacks import ModelCheckpoint  

checkpointer = ModelCheckpoint(filepath='saved_models/weights.best.Resnet50.hdf5', 
                               verbose=1, save_best_only=True)

Resnet50_model.fit(train_Resnet50, train_targets, 
          validation_data=(valid_Resnet50, valid_targets),
          epochs=25, batch_size=32, callbacks=[checkpointer], verbose=1)
Train on 6680 samples, validate on 835 samples
Epoch 1/25
6624/6680 [============================>.] - ETA: 0s - loss: 1.7753 - acc: 0.5805Epoch 00000: val_loss improved from inf to 0.85708, saving model to saved_models/weights.best.Resnet50.hdf5
6680/6680 [==============================] - 24s - loss: 1.7666 - acc: 0.5820 - val_loss: 0.8571 - val_acc: 0.7377
Epoch 2/25
6560/6680 [============================>.] - ETA: 0s - loss: 0.4664 - acc: 0.8587Epoch 00001: val_loss improved from 0.85708 to 0.71742, saving model to saved_models/weights.best.Resnet50.hdf5
6680/6680 [==============================] - 2s - loss: 0.4652 - acc: 0.8596 - val_loss: 0.7174 - val_acc: 0.7880
Epoch 3/25
6624/6680 [============================>.] - ETA: 0s - loss: 0.2597 - acc: 0.9230Epoch 00002: val_loss improved from 0.71742 to 0.68273, saving model to saved_models/weights.best.Resnet50.hdf5
6680/6680 [==============================] - 2s - loss: 0.2598 - acc: 0.9229 - val_loss: 0.6827 - val_acc: 0.7904
Epoch 4/25
6528/6680 [============================>.] - ETA: 0s - loss: 0.1679 - acc: 0.9496Epoch 00003: val_loss improved from 0.68273 to 0.63041, saving model to saved_models/weights.best.Resnet50.hdf5
6680/6680 [==============================] - 2s - loss: 0.1667 - acc: 0.9503 - val_loss: 0.6304 - val_acc: 0.8108
Epoch 5/25
6624/6680 [============================>.] - ETA: 0s - loss: 0.1111 - acc: 0.9692Epoch 00004: val_loss did not improve
6680/6680 [==============================] - 2s - loss: 0.1109 - acc: 0.9693 - val_loss: 0.6465 - val_acc: 0.8036
Epoch 6/25
6560/6680 [============================>.] - ETA: 0s - loss: 0.0738 - acc: 0.9811Epoch 00005: val_loss did not improve
6680/6680 [==============================] - 2s - loss: 0.0737 - acc: 0.9813 - val_loss: 0.6553 - val_acc: 0.8144
Epoch 7/25
6528/6680 [============================>.] - ETA: 0s - loss: 0.0546 - acc: 0.9859Epoch 00006: val_loss did not improve
6680/6680 [==============================] - 2s - loss: 0.0539 - acc: 0.9862 - val_loss: 0.6322 - val_acc: 0.8228
Epoch 8/25
6592/6680 [============================>.] - ETA: 0s - loss: 0.0388 - acc: 0.9914Epoch 00007: val_loss did not improve
6680/6680 [==============================] - 2s - loss: 0.0388 - acc: 0.9913 - val_loss: 0.6707 - val_acc: 0.8144
Epoch 9/25
6560/6680 [============================>.] - ETA: 0s - loss: 0.0271 - acc: 0.9944Epoch 00008: val_loss did not improve
6680/6680 [==============================] - 2s - loss: 0.0273 - acc: 0.9943 - val_loss: 0.6567 - val_acc: 0.8096
Epoch 10/25
6592/6680 [============================>.] - ETA: 0s - loss: 0.0216 - acc: 0.9948Epoch 00009: val_loss did not improve
6680/6680 [==============================] - 2s - loss: 0.0214 - acc: 0.9949 - val_loss: 0.6867 - val_acc: 0.8311
Epoch 11/25
6528/6680 [============================>.] - ETA: 0s - loss: 0.0161 - acc: 0.9971Epoch 00010: val_loss did not improve
6680/6680 [==============================] - 2s - loss: 0.0163 - acc: 0.9969 - val_loss: 0.7327 - val_acc: 0.8216
Epoch 12/25
6560/6680 [============================>.] - ETA: 0s - loss: 0.0108 - acc: 0.9979Epoch 00011: val_loss did not improve
6680/6680 [==============================] - 2s - loss: 0.0121 - acc: 0.9976 - val_loss: 0.7386 - val_acc: 0.8275
Epoch 13/25
6624/6680 [============================>.] - ETA: 0s - loss: 0.0104 - acc: 0.9980Epoch 00012: val_loss did not improve
6680/6680 [==============================] - 2s - loss: 0.0103 - acc: 0.9981 - val_loss: 0.7177 - val_acc: 0.8240
Epoch 14/25
6624/6680 [============================>.] - ETA: 0s - loss: 0.0083 - acc: 0.9977Epoch 00013: val_loss did not improve
6680/6680 [==============================] - 2s - loss: 0.0082 - acc: 0.9978 - val_loss: 0.7697 - val_acc: 0.8335
Epoch 15/25
6592/6680 [============================>.] - ETA: 0s - loss: 0.0076 - acc: 0.9983Epoch 00014: val_loss did not improve
6680/6680 [==============================] - 2s - loss: 0.0075 - acc: 0.9984 - val_loss: 0.8035 - val_acc: 0.8299
Epoch 16/25
6592/6680 [============================>.] - ETA: 0s - loss: 0.0075 - acc: 0.9982Epoch 00015: val_loss did not improve
6680/6680 [==============================] - 2s - loss: 0.0075 - acc: 0.9982 - val_loss: 0.8047 - val_acc: 0.8311
Epoch 17/25
6592/6680 [============================>.] - ETA: 0s - loss: 0.0062 - acc: 0.9982Epoch 00016: val_loss did not improve
6680/6680 [==============================] - 2s - loss: 0.0061 - acc: 0.9982 - val_loss: 0.8021 - val_acc: 0.8287
Epoch 18/25
6528/6680 [============================>.] - ETA: 0s - loss: 0.0055 - acc: 0.9986Epoch 00017: val_loss did not improve
6680/6680 [==============================] - 2s - loss: 0.0054 - acc: 0.9987 - val_loss: 0.8145 - val_acc: 0.8275
Epoch 19/25
6528/6680 [============================>.] - ETA: 0s - loss: 0.0054 - acc: 0.9986Epoch 00018: val_loss did not improve
6680/6680 [==============================] - 2s - loss: 0.0053 - acc: 0.9987 - val_loss: 0.8249 - val_acc: 0.8299
Epoch 20/25
6560/6680 [============================>.] - ETA: 0s - loss: 0.0049 - acc: 0.9989Epoch 00019: val_loss did not improve
6680/6680 [==============================] - 2s - loss: 0.0048 - acc: 0.9990 - val_loss: 0.8941 - val_acc: 0.8216
Epoch 21/25
6624/6680 [============================>.] - ETA: 0s - loss: 0.0053 - acc: 0.9991Epoch 00020: val_loss did not improve
6680/6680 [==============================] - 2s - loss: 0.0052 - acc: 0.9991 - val_loss: 0.9026 - val_acc: 0.8216
Epoch 22/25
6624/6680 [============================>.] - ETA: 0s - loss: 0.0049 - acc: 0.9989Epoch 00021: val_loss did not improve
6680/6680 [==============================] - 2s - loss: 0.0049 - acc: 0.9990 - val_loss: 0.8786 - val_acc: 0.8323
Epoch 23/25
6656/6680 [============================>.] - ETA: 0s - loss: 0.0044 - acc: 0.9986Epoch 00022: val_loss did not improve
6680/6680 [==============================] - 2s - loss: 0.0044 - acc: 0.9987 - val_loss: 0.8913 - val_acc: 0.8383
Epoch 24/25
6624/6680 [============================>.] - ETA: 0s - loss: 0.0034 - acc: 0.9986Epoch 00023: val_loss did not improve
6680/6680 [==============================] - 2s - loss: 0.0034 - acc: 0.9987 - val_loss: 0.9799 - val_acc: 0.8251
Epoch 25/25
6560/6680 [============================>.] - ETA: 0s - loss: 0.0048 - acc: 0.9988Epoch 00024: val_loss did not improve
6680/6680 [==============================] - 2s - loss: 0.0047 - acc: 0.9988 - val_loss: 0.9643 - val_acc: 0.8204
Out[54]:
<keras.callbacks.History at 0x7ff42dd6c7f0>

(IMPLEMENTATION) Load the Model with the Best Validation Loss

In [55]:
### TODO: Load the model weights with the best validation loss.

Resnet50_model.load_weights('saved_models/weights.best.Resnet50.hdf5')

(IMPLEMENTATION) Test the Model

Try out your model on the test dataset of dog images. Ensure that your test accuracy is greater than 60%.

In [56]:
### TODO: Calculate classification accuracy on the test dataset.

# get index of predicted dog breed for each image in test set
Resnet50_predictions = [np.argmax(Resnet50_model.predict(np.expand_dims(feature, axis=0))) for feature in test_Resnet50]

# report test accuracy
test_accuracy = 100*np.sum(np.array(Resnet50_predictions)==np.argmax(test_targets, axis=1))/len(Resnet50_predictions)
print('Test accuracy: %.4f%%' % test_accuracy)
Test accuracy: 80.6220%

(IMPLEMENTATION) Predict Dog Breed with the Model

Write a function that takes an image path as input and returns the dog breed (Affenpinscher, Afghan_hound, etc) that is predicted by your model.

Similar to the analogous function in Step 5, your function should have three steps:

  1. Extract the bottleneck features corresponding to the chosen CNN model.
  2. Supply the bottleneck features as input to the model to return the predicted vector. Note that the argmax of this prediction vector gives the index of the predicted dog breed.
  3. Use the dog_names array defined in Step 0 of this notebook to return the corresponding breed.

The functions to extract the bottleneck features can be found in extract_bottleneck_features.py, and they have been imported in an earlier code cell. To obtain the bottleneck features corresponding to your chosen CNN architecture, you need to use the function

extract_{network}

where {network}, in the above filename, should be one of VGG19, Resnet50, InceptionV3, or Xception.

In [57]:
### TODO: Write a function that takes a path to an image as input
### and returns the dog breed that is predicted by the model.

from extract_bottleneck_features import *

def Resnet50_predict_breed(img_path):
    # extract bottleneck features
    bottleneck_feature = extract_Resnet50(path_to_tensor(img_path))
    # obtain predicted vector
    predicted_vector = Resnet50_model.predict(bottleneck_feature)
    # return dog breed that is predicted by the model
    return dog_names[np.argmax(predicted_vector)]
In [58]:
import random
randomlist = random.sample(range(len(test_files)), 5)

for index, idx in enumerate(randomlist, start=1):
    print('{} > {:3d} {:25.25} {:25.25}'.format(index, idx, Resnet50_predict_breed(test_files[idx]), dog_names[np.argmax(test_targets[idx])]))
1 > 329 Golden_retriever          Golden_retriever         
2 > 115 Australian_cattle_dog     Australian_cattle_dog    
3 > 323 Cairn_terrier             Cairn_terrier            
4 >  63 Labrador_retriever        Labrador_retriever       
5 >  42 Cairn_terrier             Cairn_terrier            

Step 6: Write your Algorithm

Write an algorithm that accepts a file path to an image and first determines whether the image contains a human, dog, or neither. Then,

  • if a dog is detected in the image, return the predicted breed.
  • if a human is detected in the image, return the resembling dog breed.
  • if neither is detected in the image, provide output that indicates an error.

You are welcome to write your own functions for detecting humans and dogs in images, but feel free to use the face_detector and dog_detector functions developed above. You are required to use your CNN from Step 5 to predict dog breed.

Some sample output for our algorithm is provided below, but feel free to design your own user experience!

Sample Human Output

(IMPLEMENTATION) Write your Algorithm

In [59]:
### TODO: Write your algorithm.
### Feel free to use as many code cells as needed.

def show_img(imagefile):
    img = cv2.imread(imagefile)
    cv_rgb = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
    plt.imshow(cv_rgb)
    plt.show()

def get_dog_breed(imagefile):
    # Same as Resnet50_predict_breed, but dont wrap
    # extract bottleneck features
    bottleneck_feature = extract_Resnet50(path_to_tensor(imagefile))
    # obtain predicted vector
    predicted_vector = Resnet50_model.predict(bottleneck_feature)
    sorted_vector = predicted_vector.argsort()[::-1] 
    sorted_probs = predicted_vector[0,sorted_vector]
    sorted_names = np.array(dog_names)[sorted_vector]
    return sorted_names[:, ::-1][:, :3], sorted_probs[:, ::-1][:, :3]

def check_if_human(imagefile):
    num_faces = face_detector(imagefile)
    return num_faces > 0

def check_if_dog(imagefile):
    is_dog = dog_detector(imagefile)
    return is_dog

print('done')
done
In [60]:
human_dog_misclass=['lfw/Andrew_Fastow/Andrew_Fastow_0001.jpg',\
            'lfw/George_W_Bush/George_W_Bush_0187.jpg',\
            'lfw/Doris_Schroeder/Doris_Schroeder_0001.jpg',\
            'lfw/Roy_Williams/Roy_Williams_0001.jpg']
for idx, imgfile in enumerate(human_files_short):
    print(idx, imgfile) if imgfile in human_dog_misclass else None
print('done')
19 lfw/Andrew_Fastow/Andrew_Fastow_0001.jpg
39 lfw/George_W_Bush/George_W_Bush_0187.jpg
53 lfw/Doris_Schroeder/Doris_Schroeder_0001.jpg
69 lfw/Roy_Williams/Roy_Williams_0001.jpg
done
In [61]:
print('\n============\n')
show_img(human_files_short[75])
human = check_if_human(human_files_short[75])
print('Human? = {}'.format(human))
dog = check_if_dog(human_files_short[75])
print('Dog? = {}'.format(dog))
pred_name, pred_prob = get_dog_breed(human_files_short[75])
print(pred_name[0])
print(['%5.3f' % val for val in pred_prob[0]])

print('\n============\n')
show_img(human_files_short[17])
human = check_if_human(human_files_short[17])
print('Human? = {}'.format(human))
dog = check_if_dog(human_files_short[17])
print('Dog? = {}'.format(dog))
pred_name, pred_prob = get_dog_breed(human_files_short[17])
print(pred_name)
print(['%5.3f' % val for val in pred_prob[0]])

print('\n============\n')
show_img(human_files_short[39])
human = check_if_human(human_files_short[39])
print('Human? = {}'.format(human))
dog = check_if_dog(human_files_short[39])
print('Dog? = {}'.format(dog))
pred_name, pred_prob = get_dog_breed(human_files_short[39])
print(pred_name)
print(['%5.3f' % val for val in pred_prob[0]])

print('\n============\n')
show_img(human_files_short[19])
human = check_if_human(human_files_short[19])
print('Human? = {}'.format(human))
dog = check_if_dog(human_files_short[19])
print('Dog? = {}'.format(dog))
pred_name, pred_prob = get_dog_breed(human_files_short[19])
print(pred_name)
print(['%5.3f' % val for val in pred_prob[0]])

print('\n============\n')
show_img(dog_files_short[1])
human = check_if_human(dog_files_short[1])
print('Human? = {}'.format(human))
dog = check_if_dog(dog_files_short[1])
print('Dog? = {}'.format(dog))
pred_name, pred_prob = get_dog_breed(dog_files_short[1])
print(pred_name)
print(['%5.3f' % val for val in pred_prob[0]])

print('\n============\n')
show_img(dog_files_short[19])
human = check_if_human(dog_files_short[19])
print('Human? = {}'.format(human))
dog = check_if_dog(dog_files_short[19])
print('Dog? = {}'.format(dog))
pred_name, pred_prob = get_dog_breed(dog_files_short[19])
print(pred_name)
print(['%5.3f' % val for val in pred_prob[0]])
============

Human? = True
Dog? = False
['American_water_spaniel' 'Silky_terrier' 'Basenji']
['0.273', '0.185', '0.113']

============

Human? = True
Dog? = False
[['Xoloitzcuintli' 'Basenji' 'Italian_greyhound']]
['0.274', '0.170', '0.103']

============

Human? = True
lfw/George_W_Bush/George_W_Bush_0187.jpg
Dog? = True
[['American_water_spaniel' 'Brussels_griffon' 'Chinese_shar-pei']]
['0.275', '0.252', '0.158']

============

Human? = True
lfw/Andrew_Fastow/Andrew_Fastow_0001.jpg
Dog? = True
[['American_water_spaniel' 'Brussels_griffon' 'Dachshund']]
['0.296', '0.250', '0.133']

============

Human? = False
Dog? = True
[['Dalmatian' 'Beagle' 'Cocker_spaniel']]
['1.000', '0.000', '0.000']

============

Human? = False
Dog? = True
[['Chow_chow' 'Pekingese' 'Akita']]
['1.000', '0.000', '0.000']

Step 7: Test Your Algorithm

In this section, you will take your new algorithm for a spin! What kind of dog does the algorithm think that you look like? If you have a dog, does it predict your dog's breed accurately? If you have a cat, does it mistakenly think that your cat is a dog?

(IMPLEMENTATION) Test Your Algorithm on Sample Images!

Test your algorithm at least six images on your computer. Feel free to use any images you like. Use at least two human and two dog images.

Question 6: Is the output better than you expected :) ? Or worse :( ? Provide at least three possible points of improvement for your algorithm.

Answer:

The output is within the limits of error expected. Some humans were detected as dogs, and some dogs are misidentified as humans. However, non-human and non-dog images were correcctly identified as not containing either.

Suggested improvements:

  1. Expand the last layer to 133+1 (for human) or 133+2 ( 1 for human, the other for non-dog-non-human) and re-train the network with the corpus of dog and human images available.

     Additionally, (and this is probably a terrrible idea) we can expand the last layer to 133+ 5748 (number of individuals images) so it can identify the humans with their names as well.
  2. We havn't used image augmentation here like in the optional part. Using that, on both human and dog images and re-traininnng can improve the accuracy.

     This approach will probably negate the benifits of transfer learning..? We can train a new classifier of the type (A + B) where A = trained weights of dog images set, and B = trained weights of human images set.
  3. Algorithm tuning: For example, weight and bias initialization from a truncated gaussian distribution have been proved to be essential to the training performance in several other projects. A re-look at initializing the weights based on the following would possibly help.

     https://machinelearningmastery.com/improve-deep-learning-performance/
     http://deepdish.io/2015/02/24/network-initialization/
In [66]:
## TODO: Execute your algorithm from Step 6 on
## at least 6 images on your computer.
## Feel free to use as many code cells as needed.

def make_pred_str(nm, pr):
    pstr = 'Dog detected. Most likely breed: '
    for idx in range(1):
        pstr += str(nm[0][idx]) + ': ' + '{:.5f}'.format(pr[0][idx]) + '\t'
    return pstr +'\n'

def whosit(somepic):
    pred_str = ''
    print(somepic)
    show_img(somepic)
    human = check_if_human(somepic)
    print('Is human? = {}'.format(human))
    dog = check_if_dog(somepic)
    print('Is dog? = {}'.format(dog))
    if dog:
        pred_name, pred_prob = get_dog_breed(somepic)
        pred_str = make_pred_str(pred_name, pred_prob)
        print(pred_str)
    
    if human:
        print('Human detected.')
    
    if not (dog or human):
        print("Could not detect either human or dog...")

my_files = np.array(glob("testimages/*"))
my_files = np.sort(my_files)
#print(my_files)

for apic in my_files:
    whosit(apic)
    print(apic)
    print('============\n\n')

#print(pred_str(pred_name, pred_prob))
testimages/Alan_Greenspan_0003.jpg
Is human? = True
Is dog? = False
Human detected.
testimages/Alan_Greenspan_0003.jpg
============


testimages/Alan_Greenspan_0004.jpg
Is human? = True
Is dog? = False
Human detected.
testimages/Alan_Greenspan_0004.jpg
============


testimages/Borzoi_02154.jpg
Is human? = False
Is dog? = True
Dog detected. Most likely breed: Borzoi: 0.99980	

testimages/Borzoi_02154.jpg
============


testimages/Borzoi_02162.jpg
Is human? = False
Is dog? = True
Dog detected. Most likely breed: Irish_red_and_white_setter: 0.50630	

testimages/Borzoi_02162.jpg
============


testimages/Cartoon_1.jpg
Is human? = False
Is dog? = False
Could not detect either human or dog...
testimages/Cartoon_1.jpg
============


testimages/Cartoon_2.jpg
Is human? = False
Is dog? = False
Could not detect either human or dog...
testimages/Cartoon_2.jpg
============


testimages/Chinese_shar-pei_03556.jpg
Is human? = True
Is dog? = True
Dog detected. Most likely breed: Chinese_shar-pei: 0.99947	

Human detected.
testimages/Chinese_shar-pei_03556.jpg
============


testimages/Chinese_shar-pei_03580.jpg
Is human? = False
Is dog? = True
Dog detected. Most likely breed: Dogue_de_bordeaux: 0.91387	

testimages/Chinese_shar-pei_03580.jpg
============


testimages/Christina_Aguilera_0003.jpg
Is human? = True
Is dog? = False
Human detected.
testimages/Christina_Aguilera_0003.jpg
============


testimages/Christina_Aguilera_0004.jpg
Is human? = True
Is dog? = False
Human detected.
testimages/Christina_Aguilera_0004.jpg
============


testimages/Doris_Schroeder_0001.jpg
Is human? = True
Is dog? = True
Dog detected. Most likely breed: Silky_terrier: 0.59083	

Human detected.
testimages/Doris_Schroeder_0001.jpg
============


testimages/Doris_Schroeder_0003.jpg
Is human? = True
Is dog? = False
Human detected.
testimages/Doris_Schroeder_0003.jpg
============


testimages/George_W_Bush_0187.jpg
Is human? = True
Is dog? = True
Dog detected. Most likely breed: American_water_spaniel: 0.27501	

Human detected.
testimages/George_W_Bush_0187.jpg
============


testimages/George_W_Bush_0530.jpg
Is human? = True
Is dog? = False
Human detected.
testimages/George_W_Bush_0530.jpg
============


testimages/German_shepherd_dog_04886.jpg
Is human? = False
Is dog? = True
Dog detected. Most likely breed: German_shepherd_dog: 0.84761	

testimages/German_shepherd_dog_04886.jpg
============


testimages/German_shepherd_dog_04938.jpg
Is human? = False
Is dog? = True
Dog detected. Most likely breed: German_shepherd_dog: 0.99962	

testimages/German_shepherd_dog_04938.jpg
============


testimages/Gorilla0_from_net.jpg
Is human? = False
Is dog? = False
Could not detect either human or dog...
testimages/Gorilla0_from_net.jpg
============


testimages/Hyena2_from_net.jpeg
Is human? = False
Is dog? = False
Could not detect either human or dog...
testimages/Hyena2_from_net.jpeg
============


testimages/Hyena_from_net.jpg
Is human? = False
Is dog? = False
Could not detect either human or dog...
testimages/Hyena_from_net.jpg
============


testimages/Ibizan_hound_05679.jpg
Is human? = False
Is dog? = True
Dog detected. Most likely breed: Ibizan_hound: 0.99998	

testimages/Ibizan_hound_05679.jpg
============


testimages/Ibizan_hound_05688.jpg
Is human? = False
Is dog? = True
Dog detected. Most likely breed: Ibizan_hound: 0.97316	

testimages/Ibizan_hound_05688.jpg
============


testimages/Jelena_Dokic_0004.jpg
Is human? = True
Is dog? = False
Human detected.
testimages/Jelena_Dokic_0004.jpg
============


testimages/Jelena_Dokic_0008.jpg
Is human? = True
Is dog? = False
Human detected.
testimages/Jelena_Dokic_0008.jpg
============


testimages/Masked_ranger.jpg
Is human? = False
Is dog? = False
Could not detect either human or dog...
testimages/Masked_ranger.jpg
============


testimages/Nova_scotia_duck_tolling_retriever_07291.jpg
Is human? = False
Is dog? = True
Dog detected. Most likely breed: Nova_scotia_duck_tolling_retriever: 0.99337	

testimages/Nova_scotia_duck_tolling_retriever_07291.jpg
============


testimages/Nova_scotia_duck_tolling_retriever_07318.jpg
Is human? = False
Is dog? = True
Dog detected. Most likely breed: Nova_scotia_duck_tolling_retriever: 0.72863	

testimages/Nova_scotia_duck_tolling_retriever_07318.jpg
============


testimages/Orchid_from_net.jpg
Is human? = False
Is dog? = False
Could not detect either human or dog...
testimages/Orchid_from_net.jpg
============


testimages/Portuguese_water_dog_07971.jpg
Is human? = False
Is dog? = True
Dog detected. Most likely breed: Poodle: 0.91527	

testimages/Portuguese_water_dog_07971.jpg
============


testimages/Portuguese_water_dog_07984.jpg
Is human? = False
Is dog? = True
Dog detected. Most likely breed: Irish_water_spaniel: 0.37277	

testimages/Portuguese_water_dog_07984.jpg
============


testimages/Ttulip_from_net.jpg
Is human? = False
Is dog? = False
Could not detect either human or dog...
testimages/Ttulip_from_net.jpg
============


testimages/Will_Smith_0001.jpg
Is human? = True
Is dog? = False
Human detected.
testimages/Will_Smith_0001.jpg
============


testimages/Will_Smith_0002.jpg
Is human? = True
Is dog? = False
Human detected.
testimages/Will_Smith_0002.jpg
============


In [ ]: